diff --git a/.github/scripts/agent-guides-drive.sh b/.github/scripts/agent-guides-drive.sh
index d430d2c172..f4189a159e 100755
--- a/.github/scripts/agent-guides-drive.sh
+++ b/.github/scripts/agent-guides-drive.sh
@@ -376,7 +376,7 @@ case "$MODE" in
       hermes)   patch_hermes_tools none
                 invoke_via_connect "$OUT" -z "$PROMPT" ;;
       openclaw) patch_openclaw_agent notools
-                invoke_via_connect "$OUT" agent --local --agent ci \
+                CONNECT_CMD_OVERRIDE=openclaw invoke_via_connect "$OUT" agent --local --agent ci \
                   --model "unsloth/${UNSLOTH_MODEL_ID}" --message "$PROMPT" ;;
       *)        invoke_via_connect "$OUT" "$PROMPT" ;;
     esac
@@ -449,7 +449,7 @@ case "$MODE" in
           fi ;;
         opencode) invoke_via_connect "$out" run "$prompt" ;;
         hermes)   invoke_via_connect "$out" -z "$prompt" ;;
-        openclaw) invoke_via_connect "$out" agent --local --agent ci \
+        openclaw) CONNECT_CMD_OVERRIDE=openclaw invoke_via_connect "$out" agent --local --agent ci \
                     --model "unsloth/${UNSLOTH_MODEL_ID}" --message "$prompt" ;;
         *)        invoke_via_connect "$out" "$prompt" ;;
       esac
@@ -527,6 +527,154 @@ case "$MODE" in
     echo "[claude] attribution A/B OK (suppressed HIT, header=1 MISS)"
     ;;
 
+  # ── resume: does a launched agent's session survive exit and resume? ────
+  # Unlike the other modes, this drives the real LAUNCH path (`unsloth start
+  # <agent> ...`, the interactive default), not the --no-launch recipe. That
+  # path relocates each agent's home to a throwaway temp dir wiped on exit, so
+  # a session cannot be resumed -- unless --persist routes it to the stable
+  # Unsloth agents dir instead. We run one headless turn per pass and check
+  # whether the turn left a session in a persistent store (deterministic, no
+  # reliance on the model recalling anything), for a baseline pass and a
+  # --persist pass, and assert the expected split for this agent.
+  resume)
+    CODEWORD="PLATYPUS7"
+    T1="Remember this codeword for later: ${CODEWORD}. Reply with just the word OK."
+    T2="What codeword did I ask you to remember? Reply with just that word."
+    WORK="$WORKDIR_BASE/${AGENT}-resume"
+
+    # STABLE_HOME: the stable dir that --no-launch (and --persist) relocate to.
+    # Read it from a --no-launch probe (which also writes the agent's config
+    # there). codex/pi relocate their whole home/HOME here; opencode/claude keep
+    # their session data in a fixed user dir, so STABLE_HOME stays empty for them.
+    parse_connect
+    case "$AGENT" in
+      codex)    STABLE_HOME="$(raw_env CODEX_HOME)" ;;
+      pi)       STABLE_HOME="$(raw_env HOME)" ;;
+      *)        STABLE_HOME="" ;;
+    esac
+
+    # The persistent stores a session would land in if it were NOT wiped. We
+    # count files here before/after each turn; a positive delta means the
+    # session persisted (is resumable), zero means it went to a wiped temp dir.
+    resume_tracked_dirs() {
+      case "$AGENT" in
+        codex)    printf '%s\n' "$HOME/.codex" ;;
+        opencode) printf '%s\n' "$HOME/.local/share/opencode" "$HOME/.config/opencode" ;;
+        claude)   printf '%s\n' "$HOME/.claude" ;;
+        pi)       printf '%s\n' "$HOME/.pi" ;;
+        *)        : ;;
+      esac
+      [ -n "$STABLE_HOME" ] && printf '%s\n' "$STABLE_HOME"
+    }
+    count_session_files() {
+      local total=0 d n
+      while IFS= read -r d; do
+        [ -n "$d" ] && [ -d "$d" ] || continue
+        n="$(find "$d" -type f 2>/dev/null | wc -l)"; total=$((total + n))
+      done < <(resume_tracked_dirs)
+      echo "$total"
+    }
+
+    # The headless first-turn subcommand per agent (mirrors file-edit's map),
+    # forwarded verbatim through the launch path as passthrough args.
+    set_t1_cmd() {
+      case "$AGENT" in
+        claude)   T1_CMD=("${CLAUDE_CONNECT_FLAGS[@]}" -p "$T1") ;;
+        codex)    T1_CMD=(exec "$T1") ;;
+        opencode) T1_CMD=(run "$T1") ;;
+        pi)       T1_CMD=(-p "$T1") ;;
+        *)        guide_fail "resume mode does not cover agent '$AGENT'" ;;
+      esac
+    }
+
+    # Run one headless turn through the launch path. $1=outfile, $2="" or
+    # "--persist", rest = the agent subcommand. --yolo auto-approves so no tool
+    # prompt can hang; --api-key attaches to the already-served CI model.
+    launch_turn() {
+      local out="$1" rflag="$2"; shift 2
+      local flag=(); [ -n "$rflag" ] && flag=("$rflag")
+      run_timed "$out" unsloth start "$AGENT" "${flag[@]}" --yolo \
+        --api-key "$UNSLOTH_API_KEY" "$@"
+      local rc=$?
+      redact "$out"
+      return "$rc"
+    }
+
+    # One pass: fresh work dir, one planting turn, set RESULT to PERSISTED/WIPED
+    # from the session-store delta. Runs in the main shell (not a command
+    # substitution) so a hang's guide_fail actually fails the job and the
+    # progress lines reach the CI log. $1 = "" (baseline) or "--persist".
+    RESULT=""
+    run_pass() {
+      local rflag="$1" label="baseline"
+      [ -n "$rflag" ] && label="resume"
+      rm -rf "$WORK"; mkdir -p "$WORK"
+      set_t1_cmd
+      local out="$LOGS_DIR/${AGENT}-resume-${label}.txt"
+      local before after rc
+      before="$(count_session_files)"
+      pushd "$WORK" >/dev/null || guide_fail "could not enter work dir $WORK"
+      launch_turn "$out" "$rflag" "${T1_CMD[@]}"; rc=$?
+      popd >/dev/null || true
+      after="$(count_session_files)"
+      echo "[$AGENT] ${label}: session files ${before} -> ${after} (rc=${rc})"
+      # The turn must succeed for the delta to mean anything: an agent that writes a
+      # session file then errors would otherwise be misread as PERSISTED. Mirror the
+      # file-edit mode and fail the pass on a non-zero launch (the flagship codex recall
+      # below stays WARN-only, driven by its own launch_turn calls).
+      [ "$rc" -eq 0 ] || { echo "[$AGENT] ${label} transcript (tail):"; tail -30 "$out" 2>/dev/null || true; \
+        guide_fail "resume ${label} turn for ${AGENT} exited non-zero (rc=${rc})"; }
+      if [ "$after" -gt "$before" ]; then RESULT="PERSISTED"; else RESULT="WIPED"; fi
+    }
+
+    run_pass ""; BASELINE="$RESULT"
+    # Only the temp-dir agents (codex/pi) need the --persist pass to prove the fix.
+    # opencode/claude persist either way, so the baseline already proves it and a
+    # second full CPU turn only risks a timeout; skip it for them.
+    case "$AGENT" in
+      codex|pi) run_pass "--persist"; RESUME="$RESULT" ;;
+      *)        RESUME="n/a (persists either way)" ;;
+    esac
+
+    # Expected: codex/pi relocate their whole home to the temp dir, so a plain
+    # launch is WIPED and only --persist PERSISTS. opencode/claude keep their
+    # session data in a fixed user dir, so the baseline already PERSISTS.
+    case "$AGENT" in
+      codex|pi)        EXPECT_BASELINE="WIPED" ;;
+      opencode|claude) EXPECT_BASELINE="PERSISTED" ;;
+    esac
+
+    echo "──────────────────────────────────────────────"
+    echo "[$AGENT] RESUME EXPERIMENT"
+    echo "  baseline (unsloth start ${AGENT}):                 ${BASELINE}  (expected ${EXPECT_BASELINE})"
+    echo "  with --persist (unsloth start ${AGENT} --persist): ${RESUME}"
+    echo "──────────────────────────────────────────────"
+
+    [ "$BASELINE" = "$EXPECT_BASELINE" ] \
+      || guide_fail "baseline resume behavior for ${AGENT} was ${BASELINE}, expected ${EXPECT_BASELINE}"
+    case "$AGENT" in
+      codex|pi)
+        [ "$RESUME" = "PERSISTED" ] \
+          || guide_fail "--persist did not persist ${AGENT}'s session (got ${RESUME}); the session dir is still not stable" ;;
+    esac
+
+    # Flagship behavioral proof (codex only, WARN-only): after a --persist plant,
+    # resume the session and check the model actually recalls the codeword. A
+    # miss is not a failure (the CI model is small); the mechanism gate above is
+    # the real assertion.
+    if [ "$AGENT" = "codex" ]; then
+      rm -rf "$WORK"; mkdir -p "$WORK"
+      ( cd "$WORK" && launch_turn "$LOGS_DIR/codex-resume-plant.txt" "--persist" exec "$T1" ) || true
+      ( cd "$WORK" && launch_turn "$LOGS_DIR/codex-resume-recall.txt" "--persist" exec resume --last "$T2" ) || true
+      if grep -q "$CODEWORD" "$LOGS_DIR/codex-resume-recall.txt" 2>/dev/null; then
+        echo "[codex] behavioral recall HIT: resumed session remembered ${CODEWORD}"
+      else
+        echo "::warning::[codex] behavioral recall MISS (small CI model); mechanism gate still passed"
+      fi
+    fi
+    echo "[$AGENT] resume OK"
+    ;;
+
   *)
     echo "agent-guides-drive.sh: unknown mode '$MODE'" >&2
     exit 2
diff --git a/.github/workflows/consolidated-tests-ci.yml b/.github/workflows/consolidated-tests-ci.yml
index ae4b386589..1bb4c2bb58 100644
--- a/.github/workflows/consolidated-tests-ci.yml
+++ b/.github/workflows/consolidated-tests-ci.yml
@@ -364,7 +364,9 @@ jobs:
             tests/utils/test_attention_masks.py \
             tests/utils/test_trunc_normal_patch.py \
             tests/python/test_fast_language_model_text_only.py \
+            tests/test_bad_mappings_redirect.py \
             tests/test_prefetch_snapshot_scope.py \
+            tests/test_gemma_2b_mapper_key.py \
             --deselect 'tests/utils/test_attention_masks.py::test_run_attention_flash_varlen_receives_window_and_softcap'
           # The deselected test monkeypatches flash_attn_varlen_func, which is
           # only bound on the module when `flash_attn` is importable. flash_attn
diff --git a/.github/workflows/local-agent-guides-ci.yml b/.github/workflows/local-agent-guides-ci.yml
index 47f75dc1ba..25796bd5cf 100644
--- a/.github/workflows/local-agent-guides-ci.yml
+++ b/.github/workflows/local-agent-guides-ci.yml
@@ -471,6 +471,176 @@ jobs:
             redacted-configs/
           retention-days: 7
 
+  # ═════════════════════════════════════════════════════════════════════
+  # Job: resume
+  #   Does a conversation started with `unsloth start <agent>` survive exit
+  #   and resume? This drives the REAL launch path (not the --no-launch
+  #   recipe the other jobs use). A plain launch relocates the agent home to
+  #   a temp dir wiped on exit, so codex/pi cannot resume; --persist routes the
+  #   session to the stable Unsloth agents dir so it persists. opencode/claude
+  #   keep their session data in a fixed user dir, so they persist either way.
+  #   Dispatch-only: it is an end-to-end experiment, not a PR gate.
+  # ═════════════════════════════════════════════════════════════════════
+  resume:
+    name: resume (${{ matrix.agent }})
+    if: github.event_name == 'workflow_dispatch'
+    runs-on: ubuntu-latest
+    timeout-minutes: 60
+    strategy:
+      fail-fast: false
+      matrix:
+        # codex/pi relocate their whole home (resume broken without --persist);
+        # opencode/claude keep session data in a fixed dir (resume already works).
+        # One agent from each class proves the split end to end; openclaw/hermes
+        # share codex's relocation mechanism and are covered by the unit tests.
+        agent: [codex, opencode, claude, pi]
+    env:
+      GGUF_REPO: unsloth/gemma-4-E4B-it-GGUF
+      GGUF_FILE: gemma-4-E4B-it-UD-Q4_K_XL.gguf
+      STUDIO_PORT: '18904'
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        with:
+          persist-credentials: false
+
+      - name: Linux deps for llama.cpp prebuilt
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y --no-install-recommends \
+            libcurl4-openssl-dev libssl-dev jq
+
+      - uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e  # v6.4.0
+        with:
+          node-version: '22'
+
+      - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405  # v6.2.0
+        with:
+          python-version: '3.12'
+          cache: 'pip'
+
+      - name: Restore GGUF model file
+        id: cache-gguf
+        uses: actions/cache/restore@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
+        continue-on-error: true
+        with:
+          path: gguf-cache
+          key: ${{ runner.os }}-gguf-${{ env.GGUF_REPO }}-${{ env.GGUF_FILE }}-v1
+
+      - name: Download GGUF if cache miss
+        id: download-gguf
+        if: steps.cache-gguf.outputs.cache-hit != 'true' || steps.cache-gguf.outcome != 'success'
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        run: |
+          python -m pip install --upgrade huggingface_hub
+          mkdir -p gguf-cache
+          bash .github/scripts/hf-download-with-retry.sh "$GGUF_REPO" "$GGUF_FILE" gguf-cache
+
+      - name: Save GGUF model file
+        if: always() && steps.download-gguf.outcome == 'success'
+        uses: actions/cache/save@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
+        with:
+          path: gguf-cache
+          key: ${{ runner.os }}-gguf-${{ env.GGUF_REPO }}-${{ env.GGUF_FILE }}-v1
+
+      - name: Install Studio (--local, --no-torch)
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        run: |
+          mkdir -p logs
+          set -o pipefail
+          bash install.sh --local --no-torch 2>&1 | tee logs/install.log
+
+      - name: Serve unsloth run --disable-tools (gemma-4-E4B)
+        run: |
+          unsloth studio reset-password
+          bash .github/scripts/serve-unsloth-run.sh \
+            --gguf-file "$GITHUB_WORKSPACE/gguf-cache/${GGUF_FILE}" \
+            --port "$STUDIO_PORT" --log-dir logs \
+            --extra "--seed $UNSLOTH_SEED --temp 0" \
+            --health-timeout 900
+
+      - name: Preflight the agent's API dialect (class-a isolation)
+        env:
+          AGENT: ${{ matrix.agent }}
+        run: |
+          set -uo pipefail
+          B="$UNSLOTH_BASE_URL"; K="$UNSLOTH_API_KEY"
+          preflight_fail() {
+            echo "::error::[server/API regression] agent=$AGENT: $* (preflight failed BEFORE install/connect). Endpoint contract lives in studio/backend/routes/**.";
+            exit 1
+          }
+          code=$(curl -s -o /tmp/pf.json -w '%{http_code}' "$B/v1/models" \
+            -H "Authorization: Bearer $K") || true
+          [ "$code" = "200" ] || preflight_fail "/v1/models returned HTTP $code"
+          case "$AGENT" in
+            claude)
+              code=$(curl -s -o /tmp/pf.json -w '%{http_code}' "$B/v1/messages" \
+                -H "Authorization: Bearer $K" -H 'content-type: application/json' \
+                --max-time 120 \
+                -d "{\"model\":\"$UNSLOTH_MODEL_ID\",\"max_tokens\":16,\"messages\":[{\"role\":\"user\",\"content\":\"Hi\"}]}") || true
+              [ "$code" = "200" ] || preflight_fail "/v1/messages returned HTTP $code"
+              ;;
+            codex)
+              code=$(curl -s -o /tmp/pf.json -w '%{http_code}' "$B/v1/responses" \
+                -H "Authorization: Bearer $K" -H 'content-type: application/json' \
+                --max-time 120 \
+                -d "{\"model\":\"$UNSLOTH_MODEL_ID\",\"input\":\"Hi\",\"max_output_tokens\":16,\"stream\":true}") || true
+              [ "$code" = "200" ] || preflight_fail "/v1/responses returned HTTP $code"
+              ;;
+            *)
+              code=$(curl -s -o /tmp/pf.json -w '%{http_code}' "$B/v1/chat/completions" \
+                -H "Authorization: Bearer $K" -H 'content-type: application/json' \
+                --max-time 120 \
+                -d "{\"model\":\"$UNSLOTH_MODEL_ID\",\"max_tokens\":16,\"messages\":[{\"role\":\"user\",\"content\":\"Hi\"}]}") || true
+              [ "$code" = "200" ] || preflight_fail "/v1/chat/completions returned HTTP $code"
+              ;;
+          esac
+          echo "preflight OK for $AGENT"
+
+      - name: Install agent CLI (class-b isolation)
+        env:
+          AGENT: ${{ matrix.agent }}
+        run: bash .github/scripts/agent-guides-install.sh "$AGENT"
+
+      - name: Resume experiment (launch path)
+        env:
+          AGENT: ${{ matrix.agent }}
+        run: bash .github/scripts/agent-guides-drive.sh resume "$AGENT"
+
+      - name: Collect server logs (debug)
+        if: always()
+        run: |
+          mkdir -p logs/studio-logs
+          cp -r "$HOME/.unsloth/studio/logs/." logs/studio-logs/ 2>/dev/null || true
+          if [ -n "${UNSLOTH_API_KEY:-}" ]; then
+            grep -rlF "$UNSLOTH_API_KEY" logs redacted-configs agent-workdir 2>/dev/null | while IFS= read -r f; do
+              sed -i "s#${UNSLOTH_API_KEY}#<REDACTED>#g" "$f" 2>/dev/null || true
+            done
+          fi
+
+      - name: Stop Studio
+        if: always()
+        run: |
+          if [ -n "${UNSLOTH_SERVER_PID:-}" ] && [ "${UNSLOTH_SERVER_PID}" != "0" ]; then
+            kill "${UNSLOTH_SERVER_PID}" 2>/dev/null || true
+          fi
+          sleep 2
+          ss -tln 2>/dev/null | grep ":${STUDIO_PORT}" || true
+
+      - name: Upload logs
+        if: always()
+        continue-on-error: true
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a  # v7.0.1
+        with:
+          name: resume-${{ matrix.agent }}-log
+          path: |
+            logs/
+            agent-workdir/
+            redacted-configs/
+          retention-days: 7
+
   # ═════════════════════════════════════════════════════════════════════
   # Job 3: prompt-cache
   #   (a) curl 2-turn /v1/chat/completions: assert turn-2 cached_tokens > 0
diff --git a/.github/workflows/security-audit.yml b/.github/workflows/security-audit.yml
index 0ef2ad1e9d..1275d12216 100644
--- a/.github/workflows/security-audit.yml
+++ b/.github/workflows/security-audit.yml
@@ -2,8 +2,8 @@
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
 
 # Multi-language supply-chain audit. Triggers:
-#   - PRs touching any dependency manifest (Python / npm / Cargo) or
-#     this workflow file,
+#   - PRs touching any dependency manifest (Python / npm / Cargo), a
+#     scanner or its allowlist baseline, or this workflow file,
 #   - push to main / pip,
 #   - nightly @ 04:13 UTC so newly-published advisories surface even
 #     when no PR opens,
@@ -57,7 +57,9 @@ on:
       - 'studio/src-tauri/Cargo.lock'
       - 'pyproject.toml'
       - 'scripts/scan_packages.py'
+      - 'scripts/scan_packages_baseline.json'
       - 'scripts/scan_npm_packages.py'
+      - 'scripts/scan_npm_packages_baseline.json'
       - '.github/workflows/security-audit.yml'
   push:
     branches: [main, pip]
diff --git a/.github/workflows/studio-inference-smoke.yml b/.github/workflows/studio-inference-smoke.yml
index aebf90380a..f540c11da4 100644
--- a/.github/workflows/studio-inference-smoke.yml
+++ b/.github/workflows/studio-inference-smoke.yml
@@ -444,6 +444,8 @@ jobs:
           python - <<'PY'
           import json
           import os
+          import time
+          import urllib.error
           import urllib.request
 
           BASE = os.environ["BASE_URL"]
@@ -464,8 +466,24 @@ jobs:
                       "Content-Type": "application/json",
                   },
               )
-              with urllib.request.urlopen(req, timeout = timeout) as resp:
-                  return resp.status, json.loads(resp.read().decode())
+              # Shared CI runners stall sporadically, so retry transport-level
+              # failures only; HTTP status errors surface immediately. Bounded
+              # to fit the job's timeout-minutes: short probes get 3 full
+              # attempts, long probes one retry capped at 300s (a healthy
+              # server answers a retry quickly; a stalled one never does).
+              attempts = 3 if timeout <= 300 else 2
+              for attempt in range(attempts):
+                  try:
+                      t = timeout if attempt == 0 else min(timeout, 300)
+                      with urllib.request.urlopen(req, timeout = t) as resp:
+                          return resp.status, json.loads(resp.read().decode())
+                  except urllib.error.HTTPError:
+                      raise
+                  except (TimeoutError, ConnectionError, urllib.error.URLError) as exc:
+                      if attempt == attempts - 1:
+                          raise
+                      print(f"[retry] {path}: {exc!r}", flush = True)
+                      time.sleep(15)
 
           def post_sse(path, body, *, timeout = 600):
               """POST a streaming request and accumulate the assistant
@@ -938,6 +956,8 @@ jobs:
           import base64
           import json
           import os
+          import time
+          import urllib.error
           import urllib.request
           from openai import OpenAI
           from anthropic import Anthropic
@@ -956,8 +976,24 @@ jobs:
                       "Content-Type":  "application/json",
                   },
               )
-              with urllib.request.urlopen(req, timeout = timeout) as resp:
-                  return resp.status, json.loads(resp.read().decode())
+              # Shared CI runners stall sporadically, so retry transport-level
+              # failures only; HTTP status errors surface immediately. Bounded
+              # to fit the job's timeout-minutes: short probes get 3 full
+              # attempts, long probes one retry capped at 300s (a healthy
+              # server answers a retry quickly; a stalled one never does).
+              attempts = 3 if timeout <= 300 else 2
+              for attempt in range(attempts):
+                  try:
+                      t = timeout if attempt == 0 else min(timeout, 300)
+                      with urllib.request.urlopen(req, timeout = t) as resp:
+                          return resp.status, json.loads(resp.read().decode())
+                  except urllib.error.HTTPError:
+                      raise
+                  except (TimeoutError, ConnectionError, urllib.error.URLError) as exc:
+                      if attempt == attempts - 1:
+                          raise
+                      print(f"[retry] {path}: {exc!r}", flush = True)
+                      time.sleep(15)
 
           # ── 1. response_format = json_object (JSON mode) ─────────────
           # llama.cpp's HTTP server supports OpenAI-compatible JSON
diff --git a/.github/workflows/studio-mac-inference-smoke.yml b/.github/workflows/studio-mac-inference-smoke.yml
index d562294d42..03c0a8580d 100644
--- a/.github/workflows/studio-mac-inference-smoke.yml
+++ b/.github/workflows/studio-mac-inference-smoke.yml
@@ -430,6 +430,8 @@ jobs:
           python - <<'PY'
           import json
           import os
+          import time
+          import urllib.error
           import urllib.request
 
           BASE = os.environ["BASE_URL"]
@@ -450,8 +452,24 @@ jobs:
                       "Content-Type": "application/json",
                   },
               )
-              with urllib.request.urlopen(req, timeout = timeout) as resp:
-                  return resp.status, json.loads(resp.read().decode())
+              # Shared CI runners stall sporadically, so retry transport-level
+              # failures only; HTTP status errors surface immediately. Bounded
+              # to fit the job's timeout-minutes: short probes get 3 full
+              # attempts, long probes one retry capped at 300s (a healthy
+              # server answers a retry quickly; a stalled one never does).
+              attempts = 3 if timeout <= 300 else 2
+              for attempt in range(attempts):
+                  try:
+                      t = timeout if attempt == 0 else min(timeout, 300)
+                      with urllib.request.urlopen(req, timeout = t) as resp:
+                          return resp.status, json.loads(resp.read().decode())
+                  except urllib.error.HTTPError:
+                      raise
+                  except (TimeoutError, ConnectionError, urllib.error.URLError) as exc:
+                      if attempt == attempts - 1:
+                          raise
+                      print(f"[retry] {path}: {exc!r}", flush = True)
+                      time.sleep(15)
 
           def post_sse(path, body, *, timeout = 600):
               """POST a streaming request and accumulate the assistant
@@ -825,6 +843,8 @@ jobs:
           import base64
           import json
           import os
+          import time
+          import urllib.error
           import urllib.request
           from openai import OpenAI
           from anthropic import Anthropic
@@ -848,8 +868,24 @@ jobs:
                       "Content-Type":  "application/json",
                   },
               )
-              with urllib.request.urlopen(req, timeout = timeout) as resp:
-                  return resp.status, json.loads(resp.read().decode())
+              # Shared CI runners stall sporadically, so retry transport-level
+              # failures only; HTTP status errors surface immediately. Bounded
+              # to fit the job's timeout-minutes: short probes get 3 full
+              # attempts, long probes one retry capped at 300s (a healthy
+              # server answers a retry quickly; a stalled one never does).
+              attempts = 3 if timeout <= 300 else 2
+              for attempt in range(attempts):
+                  try:
+                      t = timeout if attempt == 0 else min(timeout, 300)
+                      with urllib.request.urlopen(req, timeout = t) as resp:
+                          return resp.status, json.loads(resp.read().decode())
+                  except urllib.error.HTTPError:
+                      raise
+                  except (TimeoutError, ConnectionError, urllib.error.URLError) as exc:
+                      if attempt == attempts - 1:
+                          raise
+                      print(f"[retry] {path}: {exc!r}", flush = True)
+                      time.sleep(15)
 
           # ── 1. response_format = json_object (JSON mode) ─────────────
           # llama.cpp's HTTP server supports OpenAI-compatible JSON
diff --git a/.github/workflows/studio-windows-inference-smoke.yml b/.github/workflows/studio-windows-inference-smoke.yml
index 0bc216d65a..0453c9212a 100644
--- a/.github/workflows/studio-windows-inference-smoke.yml
+++ b/.github/workflows/studio-windows-inference-smoke.yml
@@ -634,6 +634,8 @@ jobs:
           python - <<'PY'
           import json
           import os
+          import time
+          import urllib.error
           import urllib.request
 
           BASE = os.environ["BASE_URL"]
@@ -656,8 +658,24 @@ jobs:
                       "Content-Type": "application/json",
                   },
               )
-              with urllib.request.urlopen(req, timeout = timeout) as resp:
-                  return resp.status, json.loads(resp.read().decode())
+              # Shared CI runners stall sporadically, so retry transport-level
+              # failures only; HTTP status errors surface immediately. Bounded
+              # to fit the job's timeout-minutes: short probes get 3 full
+              # attempts, long probes one retry capped at 300s (a healthy
+              # server answers a retry quickly; a stalled one never does).
+              attempts = 3 if timeout <= 300 else 2
+              for attempt in range(attempts):
+                  try:
+                      t = timeout if attempt == 0 else min(timeout, 300)
+                      with urllib.request.urlopen(req, timeout = t) as resp:
+                          return resp.status, json.loads(resp.read().decode())
+                  except urllib.error.HTTPError:
+                      raise
+                  except (TimeoutError, ConnectionError, urllib.error.URLError) as exc:
+                      if attempt == attempts - 1:
+                          raise
+                      print(f"[retry] {path}: {exc!r}", flush = True)
+                      time.sleep(15)
 
           def post_sse(path, body, *, timeout = 600):
               body = {**body, "stream": True}
@@ -1063,6 +1081,8 @@ jobs:
           import base64
           import json
           import os
+          import time
+          import urllib.error
           import urllib.request
           from openai import OpenAI
           from anthropic import Anthropic
@@ -1082,8 +1102,24 @@ jobs:
                       "Content-Type":  "application/json",
                   },
               )
-              with urllib.request.urlopen(req, timeout = timeout) as resp:
-                  return resp.status, json.loads(resp.read().decode())
+              # Shared CI runners stall sporadically, so retry transport-level
+              # failures only; HTTP status errors surface immediately. Bounded
+              # to fit the job's timeout-minutes: short probes get 3 full
+              # attempts, long probes one retry capped at 300s (a healthy
+              # server answers a retry quickly; a stalled one never does).
+              attempts = 3 if timeout <= 300 else 2
+              for attempt in range(attempts):
+                  try:
+                      t = timeout if attempt == 0 else min(timeout, 300)
+                      with urllib.request.urlopen(req, timeout = t) as resp:
+                          return resp.status, json.loads(resp.read().decode())
+                  except urllib.error.HTTPError:
+                      raise
+                  except (TimeoutError, ConnectionError, urllib.error.URLError) as exc:
+                      if attempt == attempts - 1:
+                          raise
+                      print(f"[retry] {path}: {exc!r}", flush = True)
+                      time.sleep(15)
 
           # ── 1. response_format = json_object (JSON mode) ─────────────
           status, data = post("/v1/chat/completions", {
@@ -1334,42 +1370,75 @@ jobs:
             try { Add-MpPreference -ExclusionPath $p -ErrorAction Stop } catch { }
           }
 
-      - name: Hide Visual Studio + CMake (simulate a host with no build tools)
+      - name: Prepare no-build-tools simulation
         shell: pwsh
         run: |
           $ErrorActionPreference = 'Stop'
-          # A Program Files dir can hold a transient handle (Defender / MSBuild node)
-          # so Rename-Item intermittently fails with "Access is denied"; retry to ride it out.
-          function Rename-WithRetry($Path, $NewName) {
-            for ($i = 1; $i -le 6; $i++) {
-              try { Rename-Item -LiteralPath $Path -NewName $NewName -ErrorAction Stop; return }
-              catch { if ($i -eq 6) { throw }; Start-Sleep -Seconds 3 }
+          $root = Join-Path $env:GITHUB_WORKSPACE 'no-build-tools'
+          $pf = Join-Path $root 'ProgramFiles'
+          $pfx86 = Join-Path $root 'ProgramFilesx86'
+          New-Item -ItemType Directory -Force -Path $pf, $pfx86 | Out-Null
+
+          $blocked = [System.Collections.Generic.HashSet[string]]::new([StringComparer]::OrdinalIgnoreCase)
+          foreach ($tool in @('cmake', 'cl.exe')) {
+            foreach ($cmd in (Get-Command $tool -All -ErrorAction SilentlyContinue)) {
+              if ($cmd.Source) {
+                $dir = Split-Path -Parent $cmd.Source
+                if ($dir) {
+                  [void] $blocked.Add(
+                    [Environment]::ExpandEnvironmentVariables($dir).Trim().Trim('"').TrimEnd('\'))
+                }
+              }
             }
           }
-          # Rename the Visual Studio install roots (incl. the Installer that holds
-          # vswhere.exe) so Find-VsBuildTools' vswhere + filesystem scan both miss.
-          foreach ($d in @("$env:ProgramFiles\Microsoft Visual Studio", "${env:ProgramFiles(x86)}\Microsoft Visual Studio")) {
-            if (Test-Path -LiteralPath $d) {
-              Rename-WithRetry $d ((Split-Path $d -Leaf) + '.vsoff')
-              Write-Host "Hid VS: $d"
-            }
+          # Normalized comparison so registry spellings (trailing slash,
+          # unexpanded %VAR%) still match.
+          function Test-Blocked([string]$p) {
+            $n = [Environment]::ExpandEnvironmentVariables($p).Trim().Trim('"').TrimEnd('\')
+            return $blocked.Contains($n)
           }
-          # Surgically rename each cmake executable on PATH (not its parent dir --
-          # cmake can share a dir with other shims) so Get-Command cmake fails.
-          $hidden = @()
-          foreach ($c in (Get-Command cmake -All -ErrorAction SilentlyContinue)) {
-            if ($c.Source -and (Test-Path -LiteralPath $c.Source)) {
-              Rename-WithRetry $c.Source ((Split-Path $c.Source -Leaf) + '.off')
-              $hidden += $c.Source
-              Write-Host "Hid cmake: $($c.Source)"
-            }
+
+          $pathParts = $env:Path -split [IO.Path]::PathSeparator |
+            Where-Object { $_ -and -not (Test-Blocked $_) }
+          $noBuildToolsPath = $pathParts -join [IO.Path]::PathSeparator
+
+          # install.ps1's Refresh-SessionPath and setup.ps1's Refresh-Environment
+          # rebuild the session Path from these scopes mid-install, so filter
+          # them too. Originals are saved for the cleanup step.
+          foreach ($scope in @('Machine', 'User')) {
+            $orig = [Environment]::GetEnvironmentVariable('Path', $scope)
+            if (-not $orig) { continue }
+            Set-Content -LiteralPath (Join-Path $root "orig-path-$scope.txt") -Value $orig -NoNewline
+            $kept = ($orig -split ';' | Where-Object { $_ -and -not (Test-Blocked $_) }) -join ';'
+            [Environment]::SetEnvironmentVariable('Path', $kept, $scope)
+            Write-Host "Filtered $scope Path scope."
+          }
+
+          "NO_BUILD_TOOLS_PROGRAMFILES=$pf" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          "NO_BUILD_TOOLS_PROGRAMFILES_X86=$pfx86" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          "NO_BUILD_TOOLS_PATH<<NO_BUILD_TOOLS_PATH_EOF" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          $noBuildToolsPath | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          "NO_BUILD_TOOLS_PATH_EOF" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+
+          Write-Host "ProgramFiles simulation root: $pf"
+          Write-Host "ProgramFiles(x86) simulation root: $pfx86"
+          if ($blocked.Count -gt 0) {
+            Write-Host "Removed build-tool PATH dirs:"
+            $blocked | Sort-Object | ForEach-Object { Write-Host "  $_" }
+          } else {
+            Write-Host "No cmake or cl.exe PATH dirs found to remove."
           }
-          ("HIDDEN_CMAKE=" + ($hidden -join '|')) | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
 
       - name: Assert Visual Studio + CMake are genuinely undetectable
         shell: pwsh
         run: |
           $ErrorActionPreference = 'Stop'
+          # Set in-script: the runner does not apply step-level env keys with
+          # parentheses (`ProgramFiles(x86)`), so vswhere still found VS.
+          if (-not $env:NO_BUILD_TOOLS_PROGRAMFILES) { Write-Error "NO_BUILD_TOOLS_* env missing (Prepare step did not run?)"; exit 1 }
+          $env:ProgramFiles = $env:NO_BUILD_TOOLS_PROGRAMFILES
+          ${env:ProgramFiles(x86)} = $env:NO_BUILD_TOOLS_PROGRAMFILES_X86
+          $env:Path = $env:NO_BUILD_TOOLS_PATH
           . (Join-Path $env:GITHUB_WORKSPACE 'tests/studio_setup_ps1/Get-FunctionSource.ps1')
           $setup = Join-Path $env:GITHUB_WORKSPACE 'studio/setup.ps1'
           foreach ($fn in @('Resolve-VsGeneratorFromLabel', 'Find-VsBuildTools')) {
@@ -1394,6 +1463,10 @@ jobs:
           # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
           HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
+          # Set in-script (see the assert step); child processes inherit these.
+          $env:ProgramFiles = $env:NO_BUILD_TOOLS_PROGRAMFILES
+          ${env:ProgramFiles(x86)} = $env:NO_BUILD_TOOLS_PROGRAMFILES_X86
+          $env:Path = $env:NO_BUILD_TOOLS_PATH
           New-Item -ItemType Directory -Force -Path logs | Out-Null
           $ProgressPreference = 'SilentlyContinue'
           & ./install.ps1 --local --no-torch *>&1 | Tee-Object -FilePath logs/install.log
@@ -1480,19 +1553,19 @@ jobs:
           [ -n "$CONTENT" ] && [ "$CONTENT" != "null" ] || { echo "::error::empty completion"; exit 1; }
           echo "Inference OK without Visual Studio: $CONTENT"
 
-      - name: Restore Visual Studio + CMake
+      - name: Clean no-build-tools simulation
         if: always()
         shell: pwsh
         run: |
-          foreach ($d in @("$env:ProgramFiles\Microsoft Visual Studio", "${env:ProgramFiles(x86)}\Microsoft Visual Studio")) {
-            $off = "$d.vsoff"
-            if (Test-Path -LiteralPath $off) { Rename-Item -LiteralPath $off -NewName (Split-Path $d -Leaf); Write-Host "Restored $d" }
-          }
-          if ($env:HIDDEN_CMAKE) {
-            foreach ($src in ($env:HIDDEN_CMAKE -split '\|')) {
-              if ($src -and (Test-Path -LiteralPath "$src.off")) { Rename-Item -LiteralPath "$src.off" -NewName (Split-Path $src -Leaf) }
+          $root = Join-Path $env:GITHUB_WORKSPACE 'no-build-tools'
+          foreach ($scope in @('Machine', 'User')) {
+            $saved = Join-Path $root "orig-path-$scope.txt"
+            if (Test-Path -LiteralPath $saved) {
+              [Environment]::SetEnvironmentVariable('Path', (Get-Content -LiteralPath $saved -Raw), $scope)
+              Write-Host "Restored $scope Path scope."
             }
           }
+          Remove-Item -LiteralPath $root -Recurse -Force -ErrorAction SilentlyContinue
 
       - name: Stop Studio
         if: always()
@@ -1540,21 +1613,34 @@ jobs:
         with:
           python-version: '3.12'
 
-      - name: Hide Visual Studio
+      - name: Prepare no-build-tools simulation
         shell: pwsh
         run: |
           $ErrorActionPreference = 'Stop'
-          # Retry the rename: a Program Files dir can hold a transient handle that
-          # makes Rename-Item intermittently fail with "Access is denied".
-          function Rename-WithRetry($Path, $NewName) {
-            for ($i = 1; $i -le 6; $i++) {
-              try { Rename-Item -LiteralPath $Path -NewName $NewName -ErrorAction Stop; return }
-              catch { if ($i -eq 6) { throw }; Start-Sleep -Seconds 3 }
+          $root = Join-Path $env:GITHUB_WORKSPACE 'no-build-tools'
+          $pf = Join-Path $root 'ProgramFiles'
+          $pfx86 = Join-Path $root 'ProgramFilesx86'
+          New-Item -ItemType Directory -Force -Path $pf, $pfx86 | Out-Null
+
+          $blocked = [System.Collections.Generic.HashSet[string]]::new([StringComparer]::OrdinalIgnoreCase)
+          foreach ($tool in @('cmake', 'cl.exe')) {
+            foreach ($cmd in (Get-Command $tool -All -ErrorAction SilentlyContinue)) {
+              if ($cmd.Source) {
+                $dir = Split-Path -Parent $cmd.Source
+                if ($dir) { [void] $blocked.Add($dir) }
+              }
             }
           }
-          foreach ($d in @("$env:ProgramFiles\Microsoft Visual Studio", "${env:ProgramFiles(x86)}\Microsoft Visual Studio")) {
-            if (Test-Path -LiteralPath $d) { Rename-WithRetry $d ((Split-Path $d -Leaf) + '.vsoff'); Write-Host "Hid VS: $d" }
-          }
+
+          $pathParts = $env:Path -split [IO.Path]::PathSeparator |
+            Where-Object { $_ -and -not $blocked.Contains($_) }
+          $noBuildToolsPath = $pathParts -join [IO.Path]::PathSeparator
+
+          "NO_BUILD_TOOLS_PROGRAMFILES=$pf" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          "NO_BUILD_TOOLS_PROGRAMFILES_X86=$pfx86" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          "NO_BUILD_TOOLS_PATH<<NO_BUILD_TOOLS_PATH_EOF" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          $noBuildToolsPath | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          "NO_BUILD_TOOLS_PATH_EOF" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
 
       - name: Windows CUDA and ROCm prebuilts exist in unslothai/llama.cpp (what GPU users download, no VS)
         env:
@@ -1577,25 +1663,34 @@ jobs:
           echo "Windows CUDA and ROCm prebuilts are available -- GPU users get them without compiling."
 
       - name: The prebuilt resolver runs without Visual Studio
+        shell: pwsh
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
         run: |
+          $ErrorActionPreference = 'Stop'
+          # pwsh: bash cannot export `ProgramFiles(x86)`; set in-script so the
+          # python child inherits the overrides.
+          $env:ProgramFiles = $env:NO_BUILD_TOOLS_PROGRAMFILES
+          ${env:ProgramFiles(x86)} = $env:NO_BUILD_TOOLS_PROGRAMFILES_X86
+          $env:Path = $env:NO_BUILD_TOOLS_PATH
           # Resolver-only (no GPU on hosted runners, so the host resolves to the
           # CPU bundle). The point is that resolution needs no compiler/VS.
           python -m pip install --upgrade huggingface_hub
-          python studio/install_llama_prebuilt.py --resolve-prebuilt latest --output-format json > /tmp/resolve.json || {
-            echo "::error::resolver exited non-zero"; cat /tmp/resolve.json || true; exit 1; }
-          cat /tmp/resolve.json
-          echo "Prebuilt resolver ran with no Visual Studio present."
+          if ($LASTEXITCODE -ne 0) { Write-Host "::error::pip install huggingface_hub failed"; exit 1 }
+          python studio/install_llama_prebuilt.py --resolve-prebuilt latest --output-format json > resolve.json
+          if ($LASTEXITCODE -ne 0) {
+            Write-Host "::error::resolver exited non-zero"
+            if (Test-Path resolve.json) { Get-Content resolve.json }
+            exit 1
+          }
+          Get-Content resolve.json
+          Write-Host "Prebuilt resolver ran with no Visual Studio present."
 
-      - name: Restore Visual Studio
+      - name: Clean no-build-tools simulation
         if: always()
         shell: pwsh
         run: |
-          foreach ($d in @("$env:ProgramFiles\Microsoft Visual Studio", "${env:ProgramFiles(x86)}\Microsoft Visual Studio")) {
-            $off = "$d.vsoff"
-            if (Test-Path -LiteralPath $off) { Rename-Item -LiteralPath $off -NewName (Split-Path $d -Leaf); Write-Host "Restored $d" }
-          }
+          Remove-Item -LiteralPath (Join-Path $env:GITHUB_WORKSPACE 'no-build-tools') -Recurse -Force -ErrorAction SilentlyContinue
 
   # ── folded from studio-setup-ps1-vs2026.yml: setup.ps1 unit tests + real-VS detection + vcredist ──
   pester:
diff --git a/.github/workflows/studio-windows-update-smoke.yml b/.github/workflows/studio-windows-update-smoke.yml
index 888b3d70a3..5b92f1a3e0 100644
--- a/.github/workflows/studio-windows-update-smoke.yml
+++ b/.github/workflows/studio-windows-update-smoke.yml
@@ -6,9 +6,9 @@
 # windows-latest runner:
 #
 #   1. install.ps1 --local --no-torch installs Studio AND auto-fetches
-#      the prebuilt llama.cpp Windows binary (llama-bNNNN-bin-win-cpu-
-#      x64 from ggml-org/llama.cpp). Hitting the source-build fallback
-#      is treated as an Unsloth bug -- Studio must always pick the
+#      the prebuilt llama.cpp Windows binary (app-<tag>-windows-x64-cpu
+#      from unslothai/llama.cpp). Hitting the source-build fallback is
+#      treated as an Unsloth bug -- Studio must always pick the
 #      prebuilt on Windows.
 #   2. unsloth studio update --local is idempotent. Two consecutive
 #      runs both report "prebuilt up to date and validated", no
diff --git a/.github/workflows/version-compat-ci.yml b/.github/workflows/version-compat-ci.yml
index e492d21e99..6becccc90a 100644
--- a/.github/workflows/version-compat-ci.yml
+++ b/.github/workflows/version-compat-ci.yml
@@ -285,6 +285,92 @@ jobs:
             tests/vllm_compat/test_extended_module_imports.py \
             -v --tb=short
 
+  # Fake-CUDA GRPO/SFT/DPO patch run against REAL TRL (latest + main). Unlike
+  # the static symbol/source greps above, this drives unsloth's actual
+  # source-transform patchers (models/rl.py + rl_replacements.py) on a CPU-only
+  # runner under the tests/conftest.py spoof harness -- no GPU, no training.
+  # Catches structural TRL drift the greps miss (e.g. TRL 1.7.0's 2->3-tuple
+  # per-token-logps return, restructured PEFT ref-adapter block) by asserting
+  # the generated Unsloth trainer still satisfies the transform contracts.
+  grpo-fake-run:
+    name: GRPO fake-run (latest + main TRL, CPU spoof)
+    runs-on: ubuntu-latest
+    timeout-minutes: 18
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        with:
+          persist-credentials: false
+          path: unsloth
+      - name: Clone unsloth-zoo @ main
+        run: |
+          for attempt in 1 2 3; do
+            rm -rf "$RUNNER_TEMP/unsloth-zoo"
+            if git clone --depth=1 https://github.com/unslothai/unsloth-zoo \
+                "$RUNNER_TEMP/unsloth-zoo"; then
+              break
+            fi
+            if [ "$attempt" -eq 3 ]; then
+              echo "::error::git clone unsloth-zoo failed after 3 attempts"
+              exit 1
+            fi
+            delay=$((5 * attempt))
+            echo "::warning::clone failed (attempt $attempt/3), retrying in ${delay}s..."
+            sleep "$delay"
+          done
+      - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405  # v6.2.0
+        with:
+          python-version: '3.12'
+          cache: 'pip'
+      - name: Install CPU torch + ecosystem + TRL latest
+        run: |
+          python -m pip install --upgrade pip
+          pip install --index-url https://download.pytorch.org/whl/cpu --extra-index-url https://pypi.org/simple \
+            'torch>=2.4,<2.11' 'torchvision<0.26' 'torchcodec<0.10'
+          # Ecosystem floors unsloth needs; TRL itself is installed last so it
+          # can pull the transformers/peft it requires.
+          pip install \
+            'transformers>=4.57' 'peft>=0.18.0' 'accelerate>=1.0' 'datasets>=3.4,<5' \
+            'bitsandbytes>=0.45.5' sentencepiece protobuf safetensors numpy 'pytest>=8' \
+            'huggingface_hub>=0.34' tqdm packaging psutil triton Pillow
+          pip install --upgrade trl
+          pip install --no-deps -e "$RUNNER_TEMP/unsloth-zoo"
+          pip install --no-deps -e ./unsloth
+      - name: Fake-run vs TRL latest
+        env:
+          UNSLOTH_IS_PRESENT: '1'
+          UNSLOTH_COMPILE_DISABLE: '1'
+          # Disable dynamo/inductor at the process level, before conftest.py's early
+          # `import unsloth`, so the GRPO hot path never compiles on the GPU-less runner
+          # (defense in depth; the CPU fake-train also flips this at runtime).
+          TORCHDYNAMO_DISABLE: '1'
+          TORCH_COMPILE_DISABLE: '1'
+          PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION: python
+        run: |
+          cd unsloth
+          python -c "import trl; print('Resolved TRL', trl.__version__)"
+          PYTHONPATH=. python -m pytest \
+            tests/version_compat/test_trl_grpo_fake_run.py \
+            tests/version_compat/test_trl_fake_train_cpu.py \
+            -v --tb=short
+      # `main` is scheduled/dispatch-only so PR jobs stay fast and a bleeding-edge
+      # TRL break does not red every PR. github.event_name is valid in a step if.
+      - name: Fake-run vs TRL main (scheduled / dispatch only)
+        if: ${{ github.event_name != 'pull_request' }}
+        env:
+          UNSLOTH_IS_PRESENT: '1'
+          UNSLOTH_COMPILE_DISABLE: '1'
+          TORCHDYNAMO_DISABLE: '1'
+          TORCH_COMPILE_DISABLE: '1'
+          PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION: python
+        run: |
+          pip install --upgrade "git+https://github.com/huggingface/trl"
+          cd unsloth
+          python -c "import trl; print('Resolved TRL', trl.__version__)"
+          PYTHONPATH=. python -m pytest \
+            tests/version_compat/test_trl_grpo_fake_run.py \
+            tests/version_compat/test_trl_fake_train_cpu.py \
+            -v --tb=short
+
   # Daily-only: same suites but with --strict on importable upstream
   # tags. Schedule-only so PR jobs stay fast; cron tolerates a flake.
   daily-fresh-fetch:
diff --git a/README.md b/README.md
index e3fd4e6980..849ee2e87b 100644
--- a/README.md
+++ b/README.md
@@ -212,7 +212,7 @@ By default `unsloth studio` binds to `127.0.0.1` (this machine only). To reach i
 ```bash
 unsloth studio --secure -p 8888
 ```
-- `-H 0.0.0.0`: bind the raw port on all network interfaces, reachable from anywhere on the network. Only use this on a trusted network.
+- `-H 0.0.0.0`: bind the raw port on all network interfaces, reachable from anywhere on the network. This also starts a public Cloudflare quick tunnel by default, which publishes an internet-reachable `https://*.trycloudflare.com` URL even behind a firewall. Both the raw port and the tunnel expose Studio beyond this machine, so only use this on a network you trust; pass `--no-cloudflare` to drop the public link while keeping the network bind.
 ```bash
 unsloth studio -H 0.0.0.0 -p 8888
 ```
diff --git a/install.ps1 b/install.ps1
index 9114f80af9..0797cd3868 100644
--- a/install.ps1
+++ b/install.ps1
@@ -469,6 +469,17 @@ function Install-UnslothStudio {
         param(
             [Parameter(Mandatory = $true)][ScriptBlock]$Command
         )
+        # Installer-pinned index installs (torch) must beat an inherited uv mirror
+        # (#6898): when the command pins an index, clear every uv index env var so
+        # it wins, then restore in finally. Other installs keep the user's mirror.
+        $savedUvIndex = $null
+        if ($Command.ToString() -match '--default-index') {
+            $savedUvIndex = @{}
+            foreach ($n in 'UV_DEFAULT_INDEX', 'UV_INDEX_URL', 'UV_INDEX', 'UV_EXTRA_INDEX_URL') {
+                $savedUvIndex[$n] = [Environment]::GetEnvironmentVariable($n)
+                Remove-Item "Env:$n" -ErrorAction SilentlyContinue
+            }
+        }
         $prevEap = $ErrorActionPreference
         $ErrorActionPreference = "Continue"
         try {
@@ -488,6 +499,7 @@ function Install-UnslothStudio {
             return [int]$LASTEXITCODE
         } finally {
             $ErrorActionPreference = $prevEap
+            if ($savedUvIndex) { foreach ($n in $savedUvIndex.Keys) { if ($null -ne $savedUvIndex[$n]) { Set-Item "Env:$n" $savedUvIndex[$n] } } }
         }
     }
 
@@ -2155,7 +2167,7 @@ exit 0
         if ($SkipTorch) {
             # No-torch: install unsloth + unsloth-zoo with --no-deps, then
             # runtime deps (typer, safetensors, transformers, etc.) with --no-deps.
-            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (migrated no-torch)" { uv pip install --python $VenvPython --no-deps --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.7.1" "unsloth-zoo>=2026.7.1" }
+            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (migrated no-torch)" { uv pip install --python $VenvPython --no-deps --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.7.2" "unsloth-zoo>=2026.7.2" }
             if ($baseInstallExit -eq 0) {
                 # Resolve pydantic WITH deps so pip pins pydantic-core
                 # to the matching version (no-torch-runtime.txt below
@@ -2169,7 +2181,7 @@ exit 0
                 }
             }
         } else {
-            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (migrated)" { uv pip install --python $VenvPython --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.7.1" "unsloth-zoo>=2026.7.1" }
+            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (migrated)" { uv pip install --python $VenvPython --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.7.2" "unsloth-zoo>=2026.7.2" }
         }
         if ($baseInstallExit -ne 0) {
             Write-Host "[ERROR] Failed to install unsloth (exit code $baseInstallExit)" -ForegroundColor Red
@@ -2200,7 +2212,7 @@ exit 0
             # ABI-incompatible torchvision/torchaudio on AMD's per-arch index.
             $visionSpec = if ($ROCmGfxArch -and $torchvisionFloorMap.ContainsKey($ROCmGfxArch)) { $torchvisionFloorMap[$ROCmGfxArch] } else { "torchvision" }
             $audioSpec = if ($ROCmGfxArch -and $torchaudioFloorMap.ContainsKey($ROCmGfxArch)) { $torchaudioFloorMap[$ROCmGfxArch] } else { "torchaudio" }
-            $torchInstallExit = Invoke-InstallCommandRetry -Label "install PyTorch (AMD ROCm)" { uv pip install --python $VenvPython --force-reinstall --index-url $ROCmIndexUrl $torchSpec $visionSpec $audioSpec }
+            $torchInstallExit = Invoke-InstallCommandRetry -Label "install PyTorch (AMD ROCm)" { uv pip install --python $VenvPython --force-reinstall --default-index $ROCmIndexUrl $torchSpec $visionSpec $audioSpec }
             if ($torchInstallExit -ne 0) {
                 # Transient AMD-index failure: fall back to a CPU base so the install
                 # still completes; Studio setup retries ROCm afterwards.
@@ -2209,7 +2221,7 @@ exit 0
                 # torch (e.g. 2.10.0+rocm on gfx110X/gfx90a) that still satisfies the CPU
                 # torch>= range, so without it uv would keep the ROCm build and only swap
                 # the companions -- a mismatched venv the flavor-repair block won't fix.
-                $torchInstallExit = Invoke-InstallCommandRetry -Label "install PyTorch (CPU fallback)" { uv pip install --python $VenvPython --force-reinstall "torch>=2.4,<2.11.0" torchvision torchaudio --index-url $TorchIndexUrl }
+                $torchInstallExit = Invoke-InstallCommandRetry -Label "install PyTorch (CPU fallback)" { uv pip install --python $VenvPython --force-reinstall "torch>=2.4,<2.11.0" torchvision torchaudio --default-index $TorchIndexUrl }
                 if ($torchInstallExit -ne 0) {
                     Write-Host "[ERROR] Failed to install PyTorch (ROCm and CPU base both failed, exit code $torchInstallExit)" -ForegroundColor Red
                     return (Exit-InstallFailure "Failed to install PyTorch (exit code $torchInstallExit)" $torchInstallExit)
@@ -2223,7 +2235,7 @@ exit 0
         } else {
             Write-TauriLog "STEP" "Installing PyTorch"
             substep "installing PyTorch ($TorchIndexUrl)..."
-            $torchInstallExit = Invoke-InstallCommandRetry -Label "install PyTorch" { uv pip install --python $VenvPython "torch>=2.4,<2.11.0" torchvision torchaudio --index-url $TorchIndexUrl }
+            $torchInstallExit = Invoke-InstallCommandRetry -Label "install PyTorch" { uv pip install --python $VenvPython "torch>=2.4,<2.11.0" torchvision torchaudio --default-index $TorchIndexUrl }
             if ($torchInstallExit -ne 0) {
                 Write-Host "[ERROR] Failed to install PyTorch (exit code $torchInstallExit)" -ForegroundColor Red
                 return (Exit-InstallFailure "Failed to install PyTorch (exit code $torchInstallExit)" $torchInstallExit)
@@ -2235,7 +2247,7 @@ exit 0
         if ($SkipTorch) {
             # No-torch: install unsloth + unsloth-zoo with --no-deps, then
             # runtime deps (typer, safetensors, transformers, etc.) with --no-deps.
-            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (no-torch)" { uv pip install --python $VenvPython --no-deps --upgrade-package unsloth --upgrade-package unsloth-zoo "unsloth>=2026.7.1" "unsloth-zoo>=2026.7.1" }
+            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (no-torch)" { uv pip install --python $VenvPython --no-deps --upgrade-package unsloth --upgrade-package unsloth-zoo "unsloth>=2026.7.2" "unsloth-zoo>=2026.7.2" }
             if ($baseInstallExit -eq 0) {
                 # Same pydantic-with-deps trick as the migrated branch.
                 $baseInstallExit = Invoke-InstallCommandRetry -Label "install pydantic" { uv pip install --python $VenvPython pydantic }
@@ -2247,7 +2259,7 @@ exit 0
                 }
             }
         } elseif ($StudioLocalInstall) {
-            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (local)" { uv pip install --python $VenvPython --upgrade-package unsloth "unsloth>=2026.7.1" "unsloth-zoo>=2026.7.1" }
+            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (local)" { uv pip install --python $VenvPython --upgrade-package unsloth "unsloth>=2026.7.2" "unsloth-zoo>=2026.7.2" }
         } else {
             $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth" { uv pip install --python $VenvPython --upgrade-package unsloth -- "$PackageName" }
         }
@@ -2275,7 +2287,7 @@ exit 0
         Write-TauriLog "STEP" "Installing unsloth"
         substep "installing unsloth (this may take a few minutes)..."
         if ($StudioLocalInstall) {
-            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (auto torch backend)" { uv pip install --python $VenvPython "unsloth-zoo>=2026.7.1" "unsloth>=2026.7.1" --torch-backend=auto }
+            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (auto torch backend)" { uv pip install --python $VenvPython "unsloth-zoo>=2026.7.2" "unsloth>=2026.7.2" --torch-backend=auto }
             if ($baseInstallExit -ne 0) {
                 Write-Host "[ERROR] Failed to install unsloth (exit code $baseInstallExit)" -ForegroundColor Red
                 return (Exit-InstallFailure "Failed to install unsloth (exit code $baseInstallExit)" $baseInstallExit)
@@ -2306,7 +2318,7 @@ exit 0
     # keeps a stale torch==X+cpu against a CUDA index and setup.ps1 then loops on
     # "torch cpu != required cuXXX". Reinstall the right triplet when a GPU build is
     # expected: CUDA from $TorchIndexUrl, ROCm from $ROCmIndexUrl (repo.amd.com gfx*
-    # is a PEP 503 index uv resolves via --index-url, same URL the fresh ROCm install
+    # is a PEP 503 index uv resolves via --default-index, same URL the fresh ROCm install
     # above uses). --no-torch / CPU-only hosts (expected cpu) are no-ops.
     if (-not $SkipTorch) {
         $expectedTorchTag = Get-ExpectedTorchFlavorTag -TorchIndexUrl $TorchIndexUrl -ROCmIndexUrl $ROCmIndexUrl
@@ -2322,7 +2334,7 @@ exit 0
                     $visionSpec = if ($ROCmGfxArch -and $torchvisionFloorMap.ContainsKey($ROCmGfxArch)) { $torchvisionFloorMap[$ROCmGfxArch] } else { "torchvision" }
                     $audioSpec = if ($ROCmGfxArch -and $torchaudioFloorMap.ContainsKey($ROCmGfxArch)) { $torchaudioFloorMap[$ROCmGfxArch] } else { "torchaudio" }
                     substep "PyTorch flavor mismatch (installed $installedTorchTag, need ROCm) -- reinstalling correct build..." "Yellow"
-                    $torchFixExit = Invoke-InstallCommand { uv pip install --python $VenvPython --force-reinstall --index-url $ROCmIndexUrl $rocmSpec $visionSpec $audioSpec }
+                    $torchFixExit = Invoke-InstallCommand { uv pip install --python $VenvPython --force-reinstall --default-index $ROCmIndexUrl $rocmSpec $visionSpec $audioSpec }
                     if ($torchFixExit -ne 0) {
                         Write-Host "[ERROR] Failed to reinstall PyTorch with the correct ROCm build (exit code $torchFixExit)" -ForegroundColor Red
                         return (Exit-InstallFailure "Failed to reinstall PyTorch (ROCm) (exit code $torchFixExit)" $torchFixExit)
@@ -2331,7 +2343,7 @@ exit 0
                 } elseif ($expectedTorchTag -ne 'rocm') {
                     # CUDA: stale +cpu (or wrong cuXXX) against a CUDA index -> reinstall triplet.
                     substep "PyTorch flavor mismatch (installed $installedTorchTag, need $expectedTorchTag) -- reinstalling correct build..." "Yellow"
-                    $torchFixExit = Invoke-InstallCommand { uv pip install --python $VenvPython "torch>=2.4,<2.11.0" torchvision torchaudio --index-url $TorchIndexUrl --reinstall-package torch --reinstall-package torchvision --reinstall-package torchaudio }
+                    $torchFixExit = Invoke-InstallCommand { uv pip install --python $VenvPython "torch>=2.4,<2.11.0" torchvision torchaudio --default-index $TorchIndexUrl --reinstall-package torch --reinstall-package torchvision --reinstall-package torchaudio }
                     if ($torchFixExit -ne 0) {
                         Write-Host "[ERROR] Failed to reinstall PyTorch with the correct CUDA build (exit code $torchFixExit)" -ForegroundColor Red
                         return (Exit-InstallFailure "Failed to reinstall PyTorch ($expectedTorchTag) (exit code $torchFixExit)" $torchFixExit)
diff --git a/install.sh b/install.sh
index 796d80e401..3f4ea92387 100755
--- a/install.sh
+++ b/install.sh
@@ -159,6 +159,12 @@ run_maybe_quiet() {
 run_install_cmd() {
     _label="$1"
     shift
+    # Installer-pinned index installs (torch) must beat an inherited uv mirror
+    # (#6898): when we pass --default-index, neutralize every uv index env var so
+    # the pinned index wins. Other installs keep the user's mirror.
+    case " $* " in
+        *" --default-index "*) set -- env -u UV_DEFAULT_INDEX -u UV_INDEX_URL -u UV_INDEX -u UV_EXTRA_INDEX_URL "$@" ;;
+    esac
     if _is_verbose; then
         "$@" && return 0
         _rc=$?
@@ -2190,9 +2196,9 @@ _expected_torch_flavor_tag() {
     esac
 }
 
-# Whether index ($1) supports a plain --index-url reinstall. pytorch.org cuXXX /
+# Whether index ($1) supports a plain --default-index reinstall. pytorch.org cuXXX /
 # rocmX.Y AND the repo.amd.com gfx* indexes are all PEP 503 simple indexes that uv
-# resolves (torch + every transitive dep) via --index-url -- the same URLs the
+# resolves (torch + every transitive dep) via --default-index -- the same URLs the
 # fresh-install paths above already use -- so a stale wheel is auto-repairable.
 # Unknown/odd-mirror leaves -> no, so we warn rather than risk a wrong reinstall.
 _torch_index_repairable() {
@@ -2706,7 +2712,7 @@ if [ "$_MIGRATED" = true ]; then
         # to prevent transitive torch resolution.
         run_install_cmd_retry "install unsloth (migrated no-torch)" uv pip install --python "$_VENV_PY" --no-deps \
             --reinstall-package unsloth --reinstall-package unsloth-zoo \
-            "unsloth>=2026.7.1" "unsloth-zoo>=2026.7.1"
+            "unsloth>=2026.7.2" "unsloth-zoo>=2026.7.2"
         # Resolve pydantic WITH deps so pip pins pydantic-core to the
         # matching version (no-torch-runtime.txt below is --no-deps).
         # All transitive deps are torch-free.
@@ -2721,7 +2727,7 @@ if [ "$_MIGRATED" = true ]; then
         # overrides file, so UV_OVERRIDE is unset and this positional is the only cover.
         run_install_cmd_retry "install unsloth (migrated)" uv pip install --python "$_VENV_PY" \
             --reinstall-package unsloth --reinstall-package unsloth-zoo \
-            "unsloth>=2026.7.1" "unsloth-zoo>=2026.7.1" ${_MLX_LM_EXCLUDE_ARG:-}
+            "unsloth>=2026.7.2" "unsloth-zoo>=2026.7.2" ${_MLX_LM_EXCLUDE_ARG:-}
     fi
     if [ "$STUDIO_LOCAL_INSTALL" = true ]; then
         substep "overlaying local repo (editable)..."
@@ -2744,7 +2750,7 @@ if [ "$_MIGRATED" = true ]; then
                     substep "repairing ROCm torch (overwritten by dependency resolution)..."
                     run_install_cmd_retry "repair ROCm torch" uv pip install --python "$_VENV_PY" \
                         "$TORCH_CONSTRAINT" torchvision torchaudio \
-                        --index-url "$TORCH_INDEX_URL" \
+                        --default-index "$TORCH_INDEX_URL" \
                         --force-reinstall
                 fi
                 ;;
@@ -2870,7 +2876,7 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
                     substep "[WARN] Radeon repo lacks a compatible wheel set for this Python; falling back to ROCm index ($TORCH_INDEX_URL)" "$C_WARN"
                     run_install_cmd_retry "install PyTorch" uv pip install --python "$_VENV_PY" \
                         "$TORCH_CONSTRAINT" torchvision torchaudio \
-                        --index-url "$TORCH_INDEX_URL"
+                        --default-index "$TORCH_INDEX_URL"
                 else
                     substep "installing PyTorch from Radeon repo (${_RADEON_BASE_URL})..."
                     # Pass explicit wheel URLs so the matched trio is
@@ -2893,18 +2899,18 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
                 substep "[WARN] Radeon repo unavailable; falling back to ROCm index ($TORCH_INDEX_URL)" "$C_WARN"
                 run_install_cmd_retry "install PyTorch" uv pip install --python "$_VENV_PY" \
                     "$TORCH_CONSTRAINT" torchvision torchaudio \
-                    --index-url "$TORCH_INDEX_URL"
+                    --default-index "$TORCH_INDEX_URL"
             fi
         else
             substep "[WARN] Radeon GPU detected but could not detect full ROCm version; falling back to ROCm index" "$C_WARN"
             run_install_cmd_retry "install PyTorch" uv pip install --python "$_VENV_PY" \
                 "$TORCH_CONSTRAINT" torchvision torchaudio \
-                --index-url "$TORCH_INDEX_URL"
+                --default-index "$TORCH_INDEX_URL"
         fi
     else
         substep "installing PyTorch ($TORCH_INDEX_URL)..."
         run_install_cmd_retry "install PyTorch" uv pip install --python "$_VENV_PY" "$TORCH_CONSTRAINT" torchvision torchaudio \
-            --index-url "$TORCH_INDEX_URL"
+            --default-index "$TORCH_INDEX_URL"
     fi
     # AMD ROCm: install bitsandbytes (once, after torch, for all ROCm paths).
     # Gate on SKIP_TORCH=false so a user running with --no-torch on a ROCm
@@ -2925,7 +2931,7 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
         # runtime deps (typer, safetensors, transformers, etc.) with --no-deps.
         run_install_cmd_retry "install unsloth (no-torch)" uv pip install --python "$_VENV_PY" --no-deps \
             --upgrade-package unsloth --upgrade-package unsloth-zoo \
-            "unsloth>=2026.7.1" "unsloth-zoo>=2026.7.1"
+            "unsloth>=2026.7.2" "unsloth-zoo>=2026.7.2"
         # Same pydantic-with-deps trick as the migrated branch.
         run_install_cmd_retry "install pydantic (with deps for compatible core)" \
             uv pip install --python "$_VENV_PY" pydantic
@@ -2943,7 +2949,7 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
         fi
     elif [ "$STUDIO_LOCAL_INSTALL" = true ]; then
         run_install_cmd_retry "install unsloth (local)" uv pip install --python "$_VENV_PY" \
-            --upgrade-package unsloth "unsloth>=2026.7.1" "unsloth-zoo>=2026.7.1"
+            --upgrade-package unsloth "unsloth>=2026.7.2" "unsloth-zoo>=2026.7.2"
         substep "overlaying local repo (editable)..."
         run_install_cmd "overlay local repo" uv pip install --python "$_VENV_PY" -e "$_REPO_ROOT" --no-deps
         substep "overlaying unsloth-zoo from git main..."
@@ -2964,7 +2970,7 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
                     substep "repairing ROCm torch (overwritten by dependency resolution)..."
                     run_install_cmd_retry "repair ROCm torch" uv pip install --python "$_VENV_PY" \
                         "$TORCH_CONSTRAINT" torchvision torchaudio \
-                        --index-url "$TORCH_INDEX_URL" \
+                        --default-index "$TORCH_INDEX_URL" \
                         --force-reinstall
                 fi
                 ;;
@@ -2975,7 +2981,7 @@ else
     tauri_log "STEP" "Installing Unsloth"
     substep "installing unsloth (this may take a few minutes)..."
     if [ "$STUDIO_LOCAL_INSTALL" = true ]; then
-        run_install_cmd_retry "install unsloth (auto torch backend)" uv pip install --python "$_VENV_PY" "unsloth-zoo>=2026.7.1" "unsloth>=2026.7.1" --torch-backend=auto
+        run_install_cmd_retry "install unsloth (auto torch backend)" uv pip install --python "$_VENV_PY" "unsloth-zoo>=2026.7.2" "unsloth>=2026.7.2" --torch-backend=auto
         substep "overlaying local repo (editable)..."
         run_install_cmd "overlay local repo" uv pip install --python "$_VENV_PY" -e "$_REPO_ROOT" --no-deps
         substep "overlaying unsloth-zoo from git main..."
@@ -2999,14 +3005,14 @@ if [ "$SKIP_TORCH" = false ] && [ -n "${TORCH_INDEX_URL:-}" ]; then
         _installed_torch_ver=$("$_VENV_PY" -c "import torch; print(torch.__version__)" 2>/dev/null || true)
         _installed_torch_tag=""
         [ -n "$_installed_torch_ver" ] && _installed_torch_tag=$(_torch_flavor_tag "$_installed_torch_ver")
-        # Repair when flavor is wrong AND the index is plain --index-url reinstallable
+        # Repair when flavor is wrong AND the index is plain --default-index reinstallable
         # (cuXXX / rocmX.Y / repo.amd.com gfx*); an unknown mirror leaf -> warn only.
         if [ -n "$_installed_torch_tag" ] && [ "$_installed_torch_tag" != "$_expected_torch_tag" ] \
            && [ "$(_torch_index_repairable "$TORCH_INDEX_URL")" = "yes" ]; then
             substep "PyTorch flavor mismatch (installed $_installed_torch_tag, need $_expected_torch_tag) -- reinstalling correct build..."
             run_install_cmd "reinstall PyTorch ($_expected_torch_tag)" uv pip install --python "$_VENV_PY" \
                 "$TORCH_CONSTRAINT" torchvision torchaudio \
-                --index-url "$TORCH_INDEX_URL" \
+                --default-index "$TORCH_INDEX_URL" \
                 --reinstall-package torch --reinstall-package torchvision --reinstall-package torchaudio
             _installed_torch_ver=$("$_VENV_PY" -c "import torch; print(torch.__version__)" 2>/dev/null || true)
             _installed_torch_tag=""
@@ -3017,7 +3023,7 @@ if [ "$SKIP_TORCH" = false ] && [ -n "${TORCH_INDEX_URL:-}" ]; then
             substep "[WARN] PyTorch is CPU-only but a $_expected_torch_tag GPU build was expected for this machine." "$C_WARN"
             substep "[WARN] Training and GPU inference will run on CPU until this is fixed." "$C_WARN"
             substep "[WARN] Re-run this installer, or reinstall the GPU build manually:" "$C_WARN"
-            substep "[WARN]   uv pip install --python \"$_VENV_PY\" \"$TORCH_CONSTRAINT\" torchvision torchaudio --index-url $TORCH_INDEX_URL --reinstall-package torch --reinstall-package torchvision --reinstall-package torchaudio" "$C_WARN"
+            substep "[WARN]   uv pip install --python \"$_VENV_PY\" \"$TORCH_CONSTRAINT\" torchvision torchaudio --default-index $TORCH_INDEX_URL --reinstall-package torch --reinstall-package torchvision --reinstall-package torchaudio" "$C_WARN"
         fi
     fi
 fi
diff --git a/pyproject.toml b/pyproject.toml
index 506f1e4c75..c6d5d8c6d3 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -73,7 +73,7 @@ triton = [
 ]
 
 huggingfacenotorch = [
-    "unsloth_zoo>=2026.7.1",
+    "unsloth_zoo>=2026.7.2",
     "wheel>=0.42.0",
     "packaging",
     "numpy",
@@ -94,7 +94,7 @@ huggingfacenotorch = [
 ]
 huggingface = [
     "unsloth[huggingfacenotorch]",
-    "unsloth_zoo>=2026.7.1",
+    "unsloth_zoo>=2026.7.2",
     "torchvision",
     "unsloth[triton]",
 ]
@@ -629,7 +629,7 @@ colab-ampere-torch220 = [
     "flash-attn>=2.6.3 ; ('linux' in sys_platform)",
 ]
 colab-new = [
-    "unsloth_zoo>=2026.7.1",
+    "unsloth_zoo>=2026.7.2",
     "packaging",
     "tyro",
     "transformers>=4.51.3,!=4.52.0,!=4.52.1,!=4.52.2,!=4.52.3,!=4.53.0,!=4.54.0,!=4.55.0,!=4.55.1,!=4.57.0,!=4.57.4,!=4.57.5,!=5.0.0,!=5.1.0,<=5.5.0",
diff --git a/scripts/scan_packages_baseline.json b/scripts/scan_packages_baseline.json
index 1d34cfb66d..27fa801a2a 100644
--- a/scripts/scan_packages_baseline.json
+++ b/scripts/scan_packages_baseline.json
@@ -39,7 +39,7 @@
       "file": "botocore/utils.py",
       "check": "Reads credential paths AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Creds: L3551:     CACHE_DIR = os.path.expanduser(os.path.join('~', '.aws', 'boto', 'cache')) | L3721:         return os.path.expanduser(os.path.join('~', '.aws', 'login', 'cache'))\nNetwork: L32: from urllib.request import getproxies, proxy_bypass",
+      "evidence": "Creds: L3551:     CACHE_DIR = os.path.expanduser(os.path.join('~', '.aws', 'boto', 'cache')) | L3719:         return os.path.expanduser(os.path.join('~', '.aws', 'login', 'cache'))\nNetwork: L32: from urllib.request import getproxies, proxy_bypass",
       "evidence_hash": "2d691bc373ab872aad23c744104596ba6d0d9f3b35aa101c7edbff4429b174c1"
     },
     {
@@ -55,23 +55,23 @@
       "file": "datasets/utils/file_utils.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L443:     while True: sha256:feba37d77721aa658e1786d2e4b67de76fefe1ceeb3ce8529d361c5241778eea",
-      "evidence_hash": "2e458563dec752d0a9896c9685d368d9906867110db315ab751e3eb6ec63f51c"
+      "evidence": "L441:     while True: sha256:ce92e38c17c524815e1f9055be77235028c1e68e41b45cbfe9c8f1b867a205da",
+      "evidence_hash": "cb36281d28a975d101121c0702ee05eeee470879520d39a8be552129333f514d"
     },
     {
       "package": "datasets",
       "file": "datasets/utils/file_utils.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L441:     while True: sha256:ce92e38c17c524815e1f9055be77235028c1e68e41b45cbfe9c8f1b867a205da",
-      "evidence_hash": "cb36281d28a975d101121c0702ee05eeee470879520d39a8be552129333f514d"
+      "evidence": "L443:     while True: sha256:feba37d77721aa658e1786d2e4b67de76fefe1ceeb3ce8529d361c5241778eea",
+      "evidence_hash": "2e458563dec752d0a9896c9685d368d9906867110db315ab751e3eb6ec63f51c"
     },
     {
       "package": "diffusers",
       "file": "diffusers/utils/import_utils.py",
       "check": "Downloads and executes remote code",
       "severity": "CRITICAL",
-      "evidence": "L1015:             return importlib.import_module(\".\" + module_name, self.__name__)",
+      "evidence": "L1052:             return importlib.import_module(\".\" + module_name, self.__name__)",
       "evidence_hash": "e584ecfdb097d9482bb19cd3992813bc1a119cfd4c40af14748bafe22900d91e"
     },
     {
@@ -79,7 +79,7 @@
       "file": "diffusers/utils/testing_utils.py",
       "check": "Harvests environment variables/secrets AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Env: L233:         value = os.environ[key]\nNetwork: L688:             response = requests.get(arry, timeout=DIFFUSERS_REQUEST_TIMEOUT) | L709:     response = requests.get(url, timeout=DIFFUSERS_REQUEST_TIMEOUT) | L728:             image = PIL.Image.open(requests.get(image, stream=True, timeout=DIFFUSERS_REQUEST_TIMEOUT).raw)",
+      "evidence": "Env: L236:         value = os.environ[key]\nNetwork: L691:             response = requests.get(arry, timeout=DIFFUSERS_REQUEST_TIMEOUT) | L712:     response = requests.get(url, timeout=DIFFUSERS_REQUEST_TIMEOUT) | L731:             image = PIL.Image.open(requests.get(image, stream=True, timeout=DIFFUSERS_REQUEST_TIMEOUT).raw)",
       "evidence_hash": "671190a6106c6ee9674e5e5942dc0940e1d2f8c78d5faf674413c2345b783fd9"
     },
     {
@@ -90,12 +90,20 @@
       "evidence": "Archive: L317: a['TarFileType'] = tarfile.open(fileobj=_fileW,mode='w')\nNetwork: L330: x['SocketType'] = _socket = socket.socket()",
       "evidence_hash": "894862e547cf91b90cd6e4b495db3fb05b7490ef0d63de7e795a7e3d9447d850"
     },
+    {
+      "package": "fastapi",
+      "file": "fastapi/routing.py",
+      "check": "C2 polling/beaconing loop detected",
+      "severity": "CRITICAL",
+      "evidence": "L586:                                 while True: sha256:251135b5ebfdd1248916449f32262575e003ef64382501c65b7e4061d67bda45",
+      "evidence_hash": "365aef4449c8089753d9398417cd76ab762cef547d75db70d87bca9c0b550ab5"
+    },
     {
       "package": "fastmcp-slim",
       "file": "fastmcp/cli/apps_dev.py",
       "check": "Creates archive with sensitive data AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Archive: L1340:     with tarfile.open(fileobj=io.BytesIO(data), mode=\"r:gz\") as tar:\nNetwork: L1291:     with httpx.Client(timeout=30.0) as client: | L1305:     with httpx.Client(timeout=30.0) as client: | L1335:     with httpx.Client(timeout=30.0) as client: | L1537:         client = httpx.AsyncClient(\nL1538:             timeout=httpx.Timeout(60.0, read=None), trust_env=False\nL1539:         ) | L1701:     async with httpx.AsyncClient(trust_env=False) as client: | L1769:                     with socket.socket(family, socket.SOCK_STREAM) as s:",
+      "evidence": "Archive: L1353:     with tarfile.open(fileobj=io.BytesIO(data), mode=\"r:gz\") as tar:\nNetwork: L1304:     with httpx.Client(timeout=30.0) as client: | L1318:     with httpx.Client(timeout=30.0) as client: | L1348:     with httpx.Client(timeout=30.0) as client: | L1549:         client = httpx.AsyncClient(\nL1550:             timeout=httpx.Timeout(60.0, read=None), trust_env=False\nL1551:         ) | L1713:     async with httpx.AsyncClient(trust_env=False) as client: | L1781:                     with socket.socket(family, socket.SOCK_STREAM) as s:",
       "evidence_hash": "73a7a72013e9f800627ea07e6dbc3beeb8c905a6a5480c8fd896f0063173d25c"
     },
     {
@@ -103,8 +111,8 @@
       "file": "fastmcp/cli/apps_dev.py",
       "check": "Enumerates filesystem AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "FS: L624:     history.replaceState(null, \"\", url); sha256:fd8dbfa8af4dea2ce43f4d441f3f81239de341b76a2eb0a33c446f6757ce5f43\nNetwork: L1291:     with httpx.Client(timeout=30.0) as client: | L1305:     with httpx.Client(timeout=30.0) as client: | L1335:     with httpx.Client(timeout=30.0) as client: | L1537:         client = httpx.AsyncClient(\nL1538:             timeout=httpx.Timeout(60.0, read=None), trust_env=False\nL1539:         ) | L1701:     async with httpx.AsyncClient(trust_env=False) as client: | L1769:                     with socket.socket(family, socket.SOCK_STREAM) as s:",
-      "evidence_hash": "6ada4a9111213bdee5ea24c70a72ec4acdc8ffe0de4a01fd9835bc261ccab8f8"
+      "evidence": "FS: L637:     history.replaceState(null, \"\", url); sha256:17068ba5bfed62c3a3007ec8bf3e0ea41ef6529b9e6112064d9afb3be9231436\nNetwork: L1304:     with httpx.Client(timeout=30.0) as client: | L1318:     with httpx.Client(timeout=30.0) as client: | L1348:     with httpx.Client(timeout=30.0) as client: | L1549:         client = httpx.AsyncClient(\nL1550:             timeout=httpx.Timeout(60.0, read=None), trust_env=False\nL1551:         ) | L1713:     async with httpx.AsyncClient(trust_env=False) as client: | L1781:                     with socket.socket(family, socket.SOCK_STREAM) as s:",
+      "evidence_hash": "e5325edfada6499540e6f0c24a0868979d275522e2b6a180aa9b5dd3280681b4"
     },
     {
       "package": "fonttools",
@@ -132,19 +140,35 @@
     },
     {
       "package": "huggingface-hub",
-      "file": "huggingface_hub/hf_api.py",
+      "file": "huggingface_hub/_sandbox.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L3746:         while True: sha256:0c73ed1a7447120b112c063b14e720c6695bc11d00eb6b912cd0f10dc3e29b31",
-      "evidence_hash": "22f50b930e44146c5350bb99e6e6ebb09feea9bf1e899e407bedc4ffaf06721b"
+      "evidence": "L1179:             while True: sha256:33ceddf9e42aae207e891e97808c518e92a0b27ab60e4326256717bfb25a3a38",
+      "evidence_hash": "802fd41d8bb17bf425e99d128c0351c820103a5efb74690a4086e542a71437b8"
+    },
+    {
+      "package": "huggingface-hub",
+      "file": "huggingface_hub/_sandbox.py",
+      "check": "Writes to /tmp and executes (staged dropper)",
+      "severity": "CRITICAL",
+      "evidence": "L83: d=/tmp/.sbx-server\nL84: if command -v wget >/dev/null 2>&1; then wget -q --header \"Authorization: Bearer $SBX_DL_TOKEN\" -O \"$d\" \"$SBX_SERVER_URL\"\nL85: elif command -v curl >/dev/null 2>&1; then curl -fsSL -H \"Authorization: Bearer $SBX_DL_TOKEN\" -o \"$d\" \"$SBX_SERVER_URL\"\nL86: else cp \"$SBX_SERVER_MOUNT/sbx-server\" \"$d\"; fi\nL87: chmod +x \"$d\"",
+      "evidence_hash": "6908a3fe328fa94ee22a119998d6ad07cfa1ba4efa2628acf240f4204fd76e22"
     },
     {
       "package": "huggingface-hub",
       "file": "huggingface_hub/hf_api.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L4600:         while True: sha256:f4a851312a1832efe1b435aa1275a82184e19cc3f47e2cd244373d56c11de272",
-      "evidence_hash": "dc8fcf44788e32f42d1cc2eb0e2deb55eb2dbf2c3a55909a7d503e450f45e602"
+      "evidence": "L4677:         while True: sha256:04afb38843e4125d1476f3f04bdad0edf1f63f8d75ad49a713b13e4bc68612fb",
+      "evidence_hash": "18877a2502c862b46a5d7e33fa7c39ab4ef32da7e1b07f596fd455f4376770c6"
+    },
+    {
+      "package": "huggingface-hub",
+      "file": "huggingface_hub/hf_api.py",
+      "check": "C2 polling/beaconing loop detected",
+      "severity": "CRITICAL",
+      "evidence": "L3746:         while True: sha256:0c73ed1a7447120b112c063b14e720c6695bc11d00eb6b912cd0f10dc3e29b31",
+      "evidence_hash": "22f50b930e44146c5350bb99e6e6ebb09feea9bf1e899e407bedc4ffaf06721b"
     },
     {
       "package": "huggingface-hub",
@@ -159,8 +183,8 @@
       "file": "huggingface_hub/utils/_http.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L443:     while True: sha256:0ab4fed32d3af10f361963371f681923481377508a405b5d8770cef75f859168",
-      "evidence_hash": "1484f6b92f41c427ba8cbc7c4695a94975fea683dfa83aa510b4b0e982be4721"
+      "evidence": "L462:     while True: sha256:c75d1ee228cf7703a8c28551d649395a1f89f69a3aba69413f5bbcbd10c31958",
+      "evidence_hash": "d4d5f83fed39b87898cf776d5dad0bf1a6388a932f5fb7997d1070b50e46213e"
     },
     {
       "package": "huggingface-hub",
@@ -218,6 +242,22 @@
       "evidence": "L5: import socket sha256:915068303029fa5806199f256fb74504c65f253f9aee8ea23d8e384bb772b1c7",
       "evidence_hash": "30be130f165f418dfd37b144c5ae333de184b95f828ab8bd4010a67b84a5f814"
     },
+    {
+      "package": "multiprocess",
+      "file": "multiprocess/tests/__init__.py",
+      "check": "Reverse shell / bind shell pattern",
+      "severity": "CRITICAL",
+      "evidence": "L3355:                     os.dup2(conn.fileno(), i) | L3387:                          \"test needs os.dup2()\") | L3405:             os.dup2(fd, newfd) | L19: import socket sha256:26a745abdc7e89da28ab943394234d8ccb415e805477c3cc1f7d4766341a4c4c",
+      "evidence_hash": "a6b9bb85e9bb6682ab0dea4f95fd9266e8802f118c76d86dd87f7ab5864872cf"
+    },
+    {
+      "package": "multiprocess",
+      "file": "multiprocess/tests/__init__.py",
+      "check": "Reverse shell / bind shell pattern",
+      "severity": "CRITICAL",
+      "evidence": "L3521:                     os.dup2(conn.fileno(), i) | L3553:                          \"test needs os.dup2()\") | L3571:             os.dup2(fd, newfd) | L20: import socket sha256:07d2933301c0dbeeb6e42381687827d8dd7cfd7471986c559ca64283d5ae6e24",
+      "evidence_hash": "db1f4ca69865ec3911d7450fe11d212b817139deda21cd7a4ee32d547a8dc452"
+    },
     {
       "package": "numba",
       "file": "numba/pycc/decorators.py",
@@ -231,7 +271,7 @@
       "file": "numba/tests/support.py",
       "check": "Reverse shell / bind shell pattern",
       "severity": "CRITICAL",
-      "evidence": "L1021:         os.dup2(w, fd) | L1026:         os.dup2(save, fd)",
+      "evidence": "L1016:         os.dup2(w, fd) | L1021:         os.dup2(save, fd)",
       "evidence_hash": "fea7aa03d48bf0f4386302fa444984c4f5dfc772cfec3f1df199fd33a52eec10"
     },
     {
@@ -298,13 +338,21 @@
       "evidence": "Env: L105:     token = os.environ.get(\"AWS_BEARER_TOKEN_BEDROCK\") | L150:     environment_token = os.environ.get(\"AWS_BEARER_TOKEN_BEDROCK\")\nNetwork: L415:         http_client: httpx.Client | None = None, | L531:         http_client: httpx.Client | None = None, | L649:         http_client: httpx.AsyncClient | None = None, | L767:         http_client: httpx.AsyncClient | None = None,",
       "evidence_hash": "92dbec8ccd79c1e0bc41e93cdd0bdbb091220616c6a1352873196e9dda6bd85c"
     },
+    {
+      "package": "openai",
+      "file": "openai/resources/beta/responses/responses.py",
+      "check": "C2 polling/beaconing loop detected",
+      "severity": "CRITICAL",
+      "evidence": "L3999:         while True: sha256:df298b6eaf3416589b79f4ef283f8fb76e54d505bfda8840673f8e6419117e2e",
+      "evidence_hash": "10ce5cb5a7097fcff4042ddcfb4802edda60aa4b7b113c8b926a52ddb76f78c2"
+    },
     {
       "package": "openai",
       "file": "openai/resources/beta/threads/runs/runs.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L1074:         while True: sha256:ef6d59a4a10b73a5af491f10af2885b7a309fda9468eb0f9572d19558d3ceb9f",
-      "evidence_hash": "43c03b55fedcbc980e5e6649c3c4493729128d280cc868349ab9590908ea5f99"
+      "evidence": "L1053:         while True: sha256:973bb1aeca2e17e022872dc343a1bf5d8fe33bfa59fe01e2f8fe875522db5bce",
+      "evidence_hash": "24626e4aa53047a515ead563b42c07c43f73a2c5b82978fa59f58ffc2859e19b"
     },
     {
       "package": "openai",
@@ -319,7 +367,7 @@
       "file": "openai/resources/responses/responses.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L3803:         while True: sha256:1ce0b5a388c747945cdfda1a71b77afdfd03ae840d7aa9fa62f02eb00aa5e29f",
+      "evidence": "L3950:         while True: sha256:1ce0b5a388c747945cdfda1a71b77afdfd03ae840d7aa9fa62f02eb00aa5e29f",
       "evidence_hash": "6de300ebb5e6e17cb51c89cbcdf08515a44655182f0776f0908a9d1043ebbcd7"
     },
     {
@@ -495,16 +543,16 @@
       "file": "sklearn/datasets/_openml.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L100:             while True: sha256:270363bb66980201e477f9b94886e4023f7a3d21b5ce026b7603a8c249a50c5b",
-      "evidence_hash": "53edbe07c312d459068d38e537b5114e65685ac3d4487b0423fa4542b5df20fe"
+      "evidence": "L100:             while True: sha256:1f05a1b4fdd843b309634f583cb5e919866ef38ec5aa0b7d8a66ac8820655594",
+      "evidence_hash": "69597a64e5670a0f9a3c2aafc0bde4160f6170a9e2dc38f2c413cfa8d22ad193"
     },
     {
       "package": "scikit-learn",
       "file": "sklearn/datasets/_openml.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L100:             while True: sha256:1f05a1b4fdd843b309634f583cb5e919866ef38ec5aa0b7d8a66ac8820655594",
-      "evidence_hash": "69597a64e5670a0f9a3c2aafc0bde4160f6170a9e2dc38f2c413cfa8d22ad193"
+      "evidence": "L100:             while True: sha256:270363bb66980201e477f9b94886e4023f7a3d21b5ce026b7603a8c249a50c5b",
+      "evidence_hash": "53edbe07c312d459068d38e537b5114e65685ac3d4487b0423fa4542b5df20fe"
     },
     {
       "package": "scikit-learn",
@@ -642,6 +690,14 @@
       "evidence": "Base64: L1211:                     content = base64.b64decode(data)\nSubprocess: L2692:                             subprocess.run(\nL2693:                                 cmd.split(), capture_output=True, text=True, check=True\nL2694:                             ) | L2995:             cmd_output = subprocess.run(\nL2996:                 (\"openssl\", \"sha512\", filename), capture_output=True, text=True\nL2997:             ) | L3707:                         out = subprocess.check_output(\nL3708:                             [\"ldd\", os.path.join(search, file)]\nL3709:                         ) | L3791:             jobs.append(functools.partial(subprocess.check_call, cmd)) | L3876:                     subprocess.check_call(\nL3877:                         shlex.split(halide_cmd_gen.get_command_line())\nL3878:                     ) | L4336:                             subprocess.check_output(\nL4337:                                 cmd_parts, stderr=subprocess.STDOUT, env=os.environ\nL4338:                             ) | L4591:                         output = subprocess.check_output(\nL4592:                             cmd_parts,\nL4593:                             stderr=subprocess.STDOUT,\nL4594:                             text=True,\nL4595:                             env=os.environ,\nL4596:                         )",
       "evidence_hash": "c09774087b702a6c5d6e2e85d9239c7c241ec938fbe9c0153e8f0b5c0710389b"
     },
+    {
+      "package": "torch",
+      "file": "torch/_inductor/codecache.py",
+      "check": "base64 decode + subprocess execution (staged payload)",
+      "severity": "CRITICAL",
+      "evidence": "Base64: L1727:                     content = base64.b64decode(data)\nSubprocess: L3270:                             subprocess.run(\nL3271:                                 cmd, capture_output=True, text=True, check=True\nL3272:                             ) | L3583:             cmd_output = subprocess.run(\nL3584:                 (\"openssl\", \"sha512\", filename), capture_output=True, text=True\nL3585:             ) | L4338:                         out = subprocess.check_output(\nL4339:                             [\"ldd\", os.path.join(search, file)]\nL4340:                         ) | L4422:             jobs.append(functools.partial(subprocess.check_call, cmd)) | L4507:                     subprocess.check_call(\nL4508:                         shlex.split(halide_cmd_gen.get_command_line())\nL4509:                     ) | L4992:                             subprocess.check_output(\nL4993:                                 cmd_parts, stderr=subprocess.STDOUT, env=os.environ\nL4994:                             ) | L5247:                         output = subprocess.check_output(\nL5248:                             cmd_parts,\nL5249:                             stderr=subprocess.STDOUT,\nL5250:                             text=True,\nL5251:                             env=os.environ,\nL5252:                         )",
+      "evidence_hash": "87f77b5f51cb84fe9950fdeeb90fe8710e1b863100e90b5e2cfb228a725bee06"
+    },
     {
       "package": "torch",
       "file": "torch/ao/__init__.py",
@@ -695,7 +751,7 @@
       "file": "torch/testing/_internal/common_utils.py",
       "check": "Harvests environment variables/secrets AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Env: L4770:         env = os.environ.copy()\nNetwork: L4832:         with request.urlopen(url, timeout=15) as f1, open(path, 'wb' if binary else 'w') as f2: | L4850:     with closing(socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as sock:",
+      "evidence": "Env: L4900:         env = os.environ.copy()\nNetwork: L4962:         with request.urlopen(url, timeout=15) as f1, open(path, 'wb' if binary else 'w') as f2: | L4980:     with closing(socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as sock:",
       "evidence_hash": "704a851b9d68c9b885b9e15538bd7e96f03875503b618fe6f126c4438edd7386"
     },
     {
@@ -706,6 +762,14 @@
       "evidence": "L32: import socket sha256:89faaaa8bc908e02dad73fd59b2b481fa91189c84b39b556c2766e71d2783bf3",
       "evidence_hash": "3d23d77ace91812a07cb9508cf352185d154176e8e8c8b9b28fa92cdbcfe0d53"
     },
+    {
+      "package": "torch",
+      "file": "torch/testing/_internal/common_utils.py",
+      "check": "Reverse shell / bind shell pattern",
+      "severity": "CRITICAL",
+      "evidence": "L32: import socket sha256:ba439cbf568b194872f1d974c02b0487e51f677b67e379400522d0992600bd2d",
+      "evidence_hash": "88e98b227573997f86eedea8e885a407b0dd549d46d4a3f0b840ec5aafe66865"
+    },
     {
       "package": "torchvision",
       "file": "torchvision/datasets/utils.py",
@@ -743,8 +807,8 @@
       "file": "transformers/testing_utils.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L1663:     while True: sha256:969e911d30c37a279ad915fb8c3d2d0a3f5705a7eb82ae6e00687388b68bbe65",
-      "evidence_hash": "2aa8e94baa805d599720a16afee6f08976482e301333e619e6c343389498ad15"
+      "evidence": "L1577:     while True: sha256:2c6152f9da685f728e58d39dfc1827bc794f52606f56983bf38b5c6d0857cd5b",
+      "evidence_hash": "cdada67f3327237f00838a6750a4908dfaf76b9ab30c1352495c340d4fbd15c9"
     },
     {
       "package": "transformers",
@@ -759,15 +823,15 @@
       "file": "transformers/testing_utils.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L1577:     while True: sha256:2c6152f9da685f728e58d39dfc1827bc794f52606f56983bf38b5c6d0857cd5b",
-      "evidence_hash": "cdada67f3327237f00838a6750a4908dfaf76b9ab30c1352495c340d4fbd15c9"
+      "evidence": "L1699:     while True: sha256:969e911d30c37a279ad915fb8c3d2d0a3f5705a7eb82ae6e00687388b68bbe65",
+      "evidence_hash": "2aa8e94baa805d599720a16afee6f08976482e301333e619e6c343389498ad15"
     },
     {
       "package": "transformers",
       "file": "transformers/testing_utils.py",
       "check": "Harvests environment variables/secrets AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Env: L284:         value = os.environ[key] | L300:         value = os.environ[key] | L2129:         env = os.environ.copy() | L2251:             for k in list(os.environ.keys()):\nNetwork: L2561:     with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:",
+      "evidence": "Env: L288:         value = os.environ[key] | L304:         value = os.environ[key] | L2165:         env = os.environ.copy() | L2287:             for k in list(os.environ.keys()):\nNetwork: L2597:     with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:",
       "evidence_hash": "73ff16aee09cf163fb3a7a04dfa2cf610595bde2f19460a579397695f728e3f4"
     },
     {
@@ -799,16 +863,16 @@
       "file": "trl/extras/vllm_client.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L146:         while True: sha256:2beedc742e1f085eaa10fd3bc40be97d2331d21887ef1b9ccdfa2150a184edfe",
-      "evidence_hash": "1540dffaaa053780e953e04c11d9c6b9c74b91cb60f3e6d87451ba7fe7db46db"
+      "evidence": "L152:         while True: sha256:93e7d409e300af445376e6defbe2d0241aa19ecf63ed41b780fbb91c7d09856f",
+      "evidence_hash": "208838617172de61bca201d2a1bbeb5aa5aaa55feb1a1069cf39214673a7d6d1"
     },
     {
       "package": "trl",
       "file": "trl/extras/vllm_client.py",
       "check": "C2 polling/beaconing loop detected",
       "severity": "CRITICAL",
-      "evidence": "L152:         while True: sha256:93e7d409e300af445376e6defbe2d0241aa19ecf63ed41b780fbb91c7d09856f",
-      "evidence_hash": "208838617172de61bca201d2a1bbeb5aa5aaa55feb1a1069cf39214673a7d6d1"
+      "evidence": "L146:         while True: sha256:2beedc742e1f085eaa10fd3bc40be97d2331d21887ef1b9ccdfa2150a184edfe",
+      "evidence_hash": "1540dffaaa053780e953e04c11d9c6b9c74b91cb60f3e6d87451ba7fe7db46db"
     },
     {
       "package": "trl",
@@ -866,6 +930,14 @@
       "evidence": "Crypto: L294:     r\"|\\b(?:xprv|xpub|bc1|0x[a-fA-F0-9]{40})\\b\",\nNetwork: L53: import urllib.request | L1757:         req = urllib.request.Request(url, headers = {\"Accept\": \"application/json\"}) | L1758:         with urllib.request.urlopen(req, timeout = 30) as resp:",
       "evidence_hash": "278ff15b0b702d37d7f0b30a1e55a31bf2b11883685718a47478fbb5ce7f5212"
     },
+    {
+      "package": "unsloth-zoo",
+      "file": "scripts/scan_packages.py",
+      "check": "Writes to /tmp and executes (staged dropper)",
+      "severity": "CRITICAL",
+      "evidence": "L308:     r\"/tmp/\\S+.*(?:subprocess|os\\.system|os\\.popen|Popen|chmod.*\\+x)\", | L308:     r\"/tmp/\\S+.*(?:subprocess|os\\.system|os\\.popen|Popen|chmod.*\\+x)\", sha256:78268349021e21bedcd2eaaa5b4a71b0de1d52e023ada914dfdc09515ee1aad8",
+      "evidence_hash": "590fe1c96c442fbea5eb8642650257bc0b0199e919b9bacdb11dfa767b6fe839"
+    },
     {
       "package": "unsloth-zoo",
       "file": "tests/security/fixtures/_build.py",
@@ -919,16 +991,16 @@
       "file": "tests/test_mlx_save_export_regressions.py",
       "check": "Writes to /tmp and executes (staged dropper)",
       "severity": "CRITICAL",
-      "evidence": "L164:         temporary_location=\"/tmp/ignored\", sha256:78837e80d48e872ef191aaacfe5e1c621a98a20df486a70a41d1a932d074a5b3",
-      "evidence_hash": "dd11376e664d0d7e7f4cc4baf57eacd4b7ae7b03222dce3912ce68b63dbfca1e"
+      "evidence": "L165:         temporary_location=\"/tmp/ignored\", sha256:9f8502377b19666288b28399633dfc6740a64d0cb70ad1615e38b1269f94bf37",
+      "evidence_hash": "b7262d6e58f2ebad961dd3e64ca6c32bba356b5044d7a642d7dbd36a58cb6c81"
     },
     {
       "package": "unsloth-zoo",
       "file": "tests/test_quantize_gguf_q2_k_l.py",
       "check": "Writes to /tmp and executes (staged dropper)",
       "severity": "CRITICAL",
-      "evidence": "L67:         input_gguf=\"/tmp/in.gguf\", sha256:32532cadc357beee1009f4e86481bdbe60a0b7bf47f6bb022b05ec1b8e15aed0",
-      "evidence_hash": "49f5b67379de17178f21a9bc93b79d6b94a70ecbdd16de86574934aac30a071d"
+      "evidence": "L67:         input_gguf=\"/tmp/in.gguf\", sha256:06789b55e8f31426c233f37ff7d3729cc9e1f61c0829abd2c00c39216c63c7ad",
+      "evidence_hash": "ad4913d9099eb9b70e09d6860b242eb5f48c67e46d9bf4ae35c1c38a267d753b"
     },
     {
       "package": "unsloth-zoo",
@@ -951,7 +1023,7 @@
       "file": "unsloth_zoo/llama_cpp.py",
       "check": "Creates archive with sensitive data AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Archive: L938:         with tarfile.open(archive_path, \"r:gz\") as archive:\nNetwork: L691:             response = requests.get(url, timeout = timeout, headers = headers, stream = stream) | L1699:                     response = requests.get(\nL1700:                         LLAMA_CPP_CONVERT_FILE, timeout = (10, 120)\nL1701:                     ) | L2862:     check = requests.get(llama_cpp_chat_file, timeout = 5)",
+      "evidence": "Archive: L938:         with tarfile.open(archive_path, \"r:gz\") as archive:\nNetwork: L691:             response = requests.get(url, timeout = timeout, headers = headers, stream = stream) | L1699:                     response = requests.get(\nL1700:                         LLAMA_CPP_CONVERT_FILE, timeout = (10, 120)\nL1701:                     ) | L2873:     check = requests.get(llama_cpp_chat_file, timeout = 5)",
       "evidence_hash": "b9f3b1652349fa8ef9ac2d1715978aca1e1632165851a00a2698dd47189e410c"
     },
     {
@@ -959,7 +1031,7 @@
       "file": "unsloth_zoo/llama_cpp.py",
       "check": "Harvests environment variables/secrets AND makes network calls",
       "severity": "CRITICAL",
-      "evidence": "Env: L125: keynames = \"\\n\" + \"\\n\".join(os.environ.keys()) | L683:     token = os.environ.get(\"GH_TOKEN\") or os.environ.get(\"GITHUB_TOKEN\")\nNetwork: L691:             response = requests.get(url, timeout = timeout, headers = headers, stream = stream) | L1699:                     response = requests.get(\nL1700:                         LLAMA_CPP_CONVERT_FILE, timeout = (10, 120)\nL1701:                     ) | L2862:     check = requests.get(llama_cpp_chat_file, timeout = 5)",
+      "evidence": "Env: L125: keynames = \"\\n\" + \"\\n\".join(os.environ.keys()) | L683:     token = os.environ.get(\"GH_TOKEN\") or os.environ.get(\"GITHUB_TOKEN\")\nNetwork: L691:             response = requests.get(url, timeout = timeout, headers = headers, stream = stream) | L1699:                     response = requests.get(\nL1700:                         LLAMA_CPP_CONVERT_FILE, timeout = (10, 120)\nL1701:                     ) | L2873:     check = requests.get(llama_cpp_chat_file, timeout = 5)",
       "evidence_hash": "9cd0b1bb59c7eb1d814d7636dfd167c34f265eb7c4521a9d88b2bdcfd535b926"
     },
     {
@@ -1002,6 +1074,14 @@
       "evidence": "Obfusc: L87:     __import__(name)\nExec: L735:         exec(\"\"\"exec _code_ in _globs_, _locs_\"\"\")",
       "evidence_hash": "3cb7d8247dea7dd3d7b21ededc0181c58c50099aeb73c9138a286f3d1ad92d4f"
     },
+    {
+      "package": "cffi",
+      "file": "cffi/_cffi_gen_src.py",
+      "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
+      "severity": "HIGH",
+      "evidence": "Obfusc: L52:     compiled = compile(source=pysrc, filename=filename, mode='exec')\nExec: L53:     exec(compiled, globs, globs)",
+      "evidence_hash": "c429e4c977a61db6b7c717b5a552fce74eda622213e49eb5467a3782fd746fb9"
+    },
     {
       "package": "cffi",
       "file": "cffi/setuptools_ext.py",
@@ -1127,7 +1207,7 @@
       "file": "numba/tests/support.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L879:     __import__(modname)\nExec: L813:     eval(co, globs, ns)",
+      "evidence": "Obfusc: L874:     __import__(modname)\nExec: L808:     eval(co, globs, ns)",
       "evidence_hash": "649a7d750f903478243b0bcb9e8020521b505fc7fedc5b696ec01f4efc096109"
     },
     {
@@ -1159,7 +1239,7 @@
       "file": "numba/tests/test_np_functions.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L7106:         exec(compile(funcstr, '<string>', 'exec'), globals(), dct)\nExec: L7106:         exec(compile(funcstr, '<string>', 'exec'), globals(), dct)",
+      "evidence": "Obfusc: L7118:         exec(compile(funcstr, '<string>', 'exec'), globals(), dct)\nExec: L7118:         exec(compile(funcstr, '<string>', 'exec'), globals(), dct)",
       "evidence_hash": "9e81164131d16056fb56ad3cd11b8d129d1ff4f5855031e8b501e0335d5c14ed"
     },
     {
@@ -1175,16 +1255,16 @@
       "file": "numpy/testing/_private/utils.py",
       "check": "Anti-analysis/sandbox evasion + suspicious behavior",
       "severity": "HIGH",
-      "evidence": "Anti: L2777:             original_trace = sys.gettrace() | L2779:                 sys.settrace(None) | L2782:                 sys.settrace(original_trace)\nSubprocess: L1478:         output = subprocess.run(cmd, capture_output=True, text=True)\nExec: L1346:     exec(astr, dict) | L1632:         exec(code, globs, locs)",
-      "evidence_hash": "27468a6828101c6c026ae25aca8aa90ef485fd62b2c8f0967479edae9c965844"
+      "evidence": "Anti: L2788:             original_trace = sys.gettrace() | L2790:                 sys.settrace(None) | L2793:                 sys.settrace(original_trace)\nSubprocess: L1486:         output = subprocess.run(cmd, capture_output=True, text=True) | L2889:     res = subprocess.run(cmd, cwd=cwd, capture_output=True, text=True,\nL2890:                          errors=\"replace\", **kwargs)\nExec: L1352:     exec(astr, dict) | L1640:         exec(code, globs, locs)",
+      "evidence_hash": "9c6961817e5b1751e572dfe0858286703bb835870ecdfd6a7a9fdd8372a5dd2b"
     },
     {
       "package": "numpy",
       "file": "numpy/testing/_private/utils.py",
       "check": "Anti-analysis/sandbox evasion + suspicious behavior",
       "severity": "HIGH",
-      "evidence": "Anti: L2788:             original_trace = sys.gettrace() | L2790:                 sys.settrace(None) | L2793:                 sys.settrace(original_trace)\nSubprocess: L1486:         output = subprocess.run(cmd, capture_output=True, text=True) | L2889:     res = subprocess.run(cmd, cwd=cwd, capture_output=True, text=True,\nL2890:                          errors=\"replace\", **kwargs)\nExec: L1352:     exec(astr, dict) | L1640:         exec(code, globs, locs)",
-      "evidence_hash": "9c6961817e5b1751e572dfe0858286703bb835870ecdfd6a7a9fdd8372a5dd2b"
+      "evidence": "Anti: L2777:             original_trace = sys.gettrace() | L2779:                 sys.settrace(None) | L2782:                 sys.settrace(original_trace)\nSubprocess: L1478:         output = subprocess.run(cmd, capture_output=True, text=True)\nExec: L1346:     exec(astr, dict) | L1632:         exec(code, globs, locs)",
+      "evidence_hash": "27468a6828101c6c026ae25aca8aa90ef485fd62b2c8f0967479edae9c965844"
     },
     {
       "package": "numpy",
@@ -1199,7 +1279,7 @@
       "file": "PIL/Image.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L422:         __import__(f\"{__spec__.parent}.{plugin}\", globals(), locals(), []) | L490:             __import__(f\"{__spec__.parent}.{plugin}\", globals(), locals(), [])\nExec: L3772: def eval(image: Image, *args: Callable[[int], float]) -> Image:",
+      "evidence": "Obfusc: L422:         __import__(f\"{__spec__.parent}.{plugin}\", globals(), locals(), []) | L490:             __import__(f\"{__spec__.parent}.{plugin}\", globals(), locals(), [])\nExec: L3776: def eval(image: Image, *args: Callable[[int], float]) -> Image:",
       "evidence_hash": "c2c1e7ae44e15862caf8de549d09db7b35e93282450f07ef61aaf5450a408c13"
     },
     {
@@ -1255,7 +1335,7 @@
       "file": "setuptools/_distutils/compilers/C/base.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L1286:         __import__(module_name)\nExec: L1113:         if lib_type not in eval(expected):",
+      "evidence": "Obfusc: L1287:         __import__(module_name)\nExec: L1114:         if lib_type not in eval(expected):",
       "evidence_hash": "368651e9818ed2d1bb009027d3bcfbf94ae30639c0882a6c2bddde97b8c4f1e5"
     },
     {
@@ -1271,7 +1351,7 @@
       "file": "setuptools/tests/config/test_pyprojecttoml.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L364:         \"setup.py\": \"__import__('setuptools').setup(include_package_data=False)\",\nExec: L98:             \"__main__.py\": \"def exec(): print('hello')\",",
+      "evidence": "Obfusc: L387:         \"setup.py\": \"__import__('setuptools').setup(include_package_data=False)\",\nExec: L98:             \"__main__.py\": \"def exec(): print('hello')\",",
       "evidence_hash": "067d41014f72a61d8b4adf25f3659d1f66a0e909f732223f48837aa7684df4e6"
     },
     {
@@ -1279,7 +1359,7 @@
       "file": "setuptools/tests/test_editable_install.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L120: SETUP_SCRIPT_STUB = \"__import__('setuptools').setup()\"\nExec: L449:         exec(finder, loc, loc)",
+      "evidence": "Obfusc: L120: SETUP_SCRIPT_STUB = \"__import__('setuptools').setup()\"\nExec: L447:         exec(finder, loc, loc)",
       "evidence_hash": "a78d7f5af7eb4ba92656cda258c195b92f6337c585c97d0823e47a9d4a2eb15d"
     },
     {
@@ -1322,12 +1402,20 @@
       "evidence": "Obfusc: L919:     c = compile(funcstr, filename, 'exec')\nExec: L163:             module = eval(import_command) | L170:                 exec(import_command, {}, namespace) | L903:                     exec(ln, {}, namespace) | L909:                     exec(ln, {}, namespace) | L920:     exec(c, namespace, funclocals)",
       "evidence_hash": "ab4f5819576a70038301668b8f3e4a781c4b757b146117d5d93eab1896a5a6cd"
     },
+    {
+      "package": "tensorboard",
+      "file": "tensorboard/plugins/projector/tf_projector_plugin/projector_binary.js",
+      "check": "Python wheel ships large JS bundle (uncommon; manually review)",
+      "severity": "HIGH",
+      "evidence": "sha256: 53c38430766be25dc672a30846ac3b9eba86aee35eb0746785ec012647c7d9a2",
+      "evidence_hash": "2c6384e8115a6d5dacf1f84d8f724832d8dc59feb442bb98ffae0857c0ccb381"
+    },
     {
       "package": "torch",
       "file": "torch/_dynamo/bytecode_debugger.py",
       "check": "Anti-analysis/sandbox evasion + suspicious behavior",
       "severity": "HIGH",
-      "evidence": "Anti: L1048:             self._old_trace = sys.gettrace() | L1049:             sys.settrace(self._settrace_callback) | L1106:             sys.settrace(self._old_trace)\nExec: L683:                         result = eval(arg, frame_globals, eval_locals) | L708:                     result = eval(cmd, frame_globals, eval_locals) | L716:                         exec(cmd, frame_globals, eval_locals)",
+      "evidence": "Anti: L1052:             self._old_trace = sys.gettrace() | L1053:             sys.settrace(self._settrace_callback) | L1113:             sys.settrace(self._old_trace)\nExec: L684:                         result = eval(arg, frame_globals, eval_locals) | L709:                     result = eval(cmd, frame_globals, eval_locals) | L717:                         exec(cmd, frame_globals, eval_locals)",
       "evidence_hash": "dc2afd1769d357c15b69802bd2799fafa059c0b1dcdd4937528fb5b601962f1b"
     },
     {
@@ -1343,7 +1431,7 @@
       "file": "torch/fx/experimental/rewriter.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L46:         code = compile(dest_ast, \"\", \"exec\")\nExec: L49:         exec(code, globals_dict)",
+      "evidence": "Obfusc: L44:         code = compile(dest_ast, \"\", \"exec\")\nExec: L47:         exec(code, globals_dict)",
       "evidence_hash": "76374f96feed416eec390458843621f33524cfb8d93ef0f3eb4cb1b47d0ad748"
     },
     {
@@ -1359,7 +1447,7 @@
       "file": "torch/package/package_importer.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L602:     def __import__(self, name, globals=None, locals=None, fromlist=(), level=0):\nExec: L412:             exec(code, ns)",
+      "evidence": "Obfusc: L599:     def __import__(self, name, globals=None, locals=None, fromlist=(), level=0):\nExec: L412:             exec(code, ns)",
       "evidence_hash": "c7c0650f0c74a086d224112f77ee76634b8f47afc047ce27fee8c7fc45560512"
     },
     {
@@ -1391,7 +1479,7 @@
       "file": "tests/test_mlx_trainer_internals.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L430:     assert ppl == pytest.approx(__import__(\"math\").exp(2.5))\nExec: L408:         def eval(self):",
+      "evidence": "Obfusc: L1158:     assert ppl == pytest.approx(__import__(\"math\").exp(2.5))\nExec: L1136:         def eval(self):",
       "evidence_hash": "c409327ef6420cc0c7224506fcb82b11bbc9838a6f2f97c9c2cfc00a40c4cdbf"
     },
     {
@@ -1407,7 +1495,7 @@
       "file": "unsloth_zoo/compiler.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L1013:             _mod = __import__(model_location, fromlist=items) | L4291:         f\"   {chr(92)}{chr(92)}   /|    Num examples = {num_examples:,} | Num Epochs = {num_train_epochs:,} | Total steps = {max_steps:,}\\\\n\"\\\\\nL4292:         f\"O^O/ {chr(92)}_/ {chr(92)}    Batch size per device = {self._train_batch_size:,} | Gradient accumulation steps = {args.gradient_accumulation_steps}\\\\n\"\\\\\nL4293:         f\"{chr(92)}        /    Data Parallel GPUs = {args.world_size} | Total batch size ({self._train_batch_size} x {args.gradient_accumulation_steps} x {args.world_size}) = {total_train_batch_size: sha256:b7ea9cdbe360ad323911014caf12a9d4ec87cb36195a511080e4e484c2fb80d2\nL4294:         f' \"-____-\"     Trainable parameters = {get_model_param_count(model, trainable_only=True):,} of {get_model_param_count(model):,} ({get_model_param_count(model, trainable_only=True)/get_model_p sha256:9e832c1e7b2815aa44dc42638d821cb9df22550db88d85bd4bfb29c13f984b53 | L4292:         f\"O^O/ {chr(92)}_/ {chr(92)}    Batch size per device = {self._train_batch_size:,} | Gradient accumulation steps = {args.gradient_accumulation_steps}\\\\n\"\\\\\nL4293:         f\"{chr(92)}        /    Data Parallel GPUs = {args.world_size} | Total batch size ({self._train_batch_size} x {args.gradient_accumulation_steps} x {args.world_size}) = {total_train_batch_size: sha256:b7ea9cdbe360ad323911014caf12a9d4ec87cb36195a511080e4e484c2fb80d2\nL4294:         f' \"-____-\"     Trainable parameters = {get_model_param_count(model, trainable_only=True):,} of {get_model_param_count(model):,} ({get_model_param_count(model, trainable_only=True)/get_model_p sha256:9e832c1e7b2815aa44dc42638d821cb9df22550db88d85bd4bfb29c13f984b53 | L4291:         f\"   {chr(92)}{chr(92)}   /|    Num examples = {num_examples:,} | Num Epochs = {num_train_epochs:,} | Total steps = {max_steps:,}\\\\n\"\\\\\nL4292:         f\"O^O/ {chr(92)}_/ {chr(92)}    Batch size per device = {self._train_batch_size:,} | Gradient accumulation steps = {args.gradient_accumulation_steps}\\\\n\"\\\\\nL4293:         f\"{chr(92)}        /    Data Parallel GPUs = {args.world_size} | Total batch size ({self._train_batch_size} x {args.gradient_accumulation_steps} x {args.world_size}) = {total_train_batch_size: sha256:b7ea9cdbe360ad323911014caf12a9d4ec87cb36195a511080e4e484c2fb80d2\nExec: L612:                 if eval(_dtype) is not None: | L613:                     dtype = eval(_dtype) | L955:         _modeling_file = eval(model_location) | L1255:     f = eval(f\"{model_location}.{module}\") | L1563:         exec(f\"def raise_{j}(*args, **kwargs): print('{function}')\", globals(), locals()) | L1564:         try: exec(f\"EMPTY_LOGITS.{function} = raise_{j}\", globals(), locals()) | L2699:         exec(f\"import {parent}\", locals(), globals()) | L2830:                 dir(eval(parent)), | L2834:             exec(f\"{parent}.{child}.forward = forward\", globals(), locals()) | L2908:         module = eval(f\"modeling_file.{module}\") | L2935:             inner_class = eval(f\"modeling_file.{inner_class}\") | L3065:                 exec(f\"from timm.layers.norm_act import {norm}\") | L3073:             forward = eval(norm).forward | L3079:             exec(f\"timm.layers.norm_act.{norm}.forward = forward\") | L3096:                 exec(f\"from timm.models._efficientnet_blocks import {block}\") | L3104:             forward = eval(block).forward | L3110:             exec(f\"timm.models._efficientnet_blocks.{block}.forward = forward\") | L3385:         exec(f\"import {model_location}\", globals()) | L3388:     modeling_file = eval(model_location) | L3401:     exec(\nL3402:         \"model_logger.addFilter(HideLoggingMessage('`use_cache`'))\", globals(), locals()\nL3403:     ) | L3405:     exec(\nL3406:         \"model_logger.addFilter(HideLoggingMessage('compile_config'))\",\nL3407:         globals(),\nL3408:         locals(),\nL3409:     ) | L3560:         source = eval(f\"modeling_file.{module}\") | L3574:         source = eval(f\"modeling_file.{module}\") | L3675:         source = eval(f\"modeling_file.{module}\") | L3713:             source = eval(f\"{model_location}.{module}\") | L3784:                 source = eval(f\"{model_location}.{module}\") | L3832:             source = eval(f\"{model_location}.{module}\") | L4054:         source = eval(f\"{model_location}.{module}\") | L4065:         exec(\nL4066:             f\"{model_location}.{module}._update_causal_mask = no_update_causal_mask\",\nL4067:             globals(),\nL4068:         ) | L4131:                 source = eval(f\"{model_location}.{module}\") | L4172:                 module_cls = eval(f\"{model_location}.{module}\") | L4209:                 module_cls = eval(f\"{model_location}.{module}\") | L4276:     exec(\nL4277:         \"from transformers.trainer import (\" + \", \".join(x for x in good_items) + \")\",\nL4278:         globals(),\nL4279:     ) | L4341:     exec(inner_training_loop, globals()) | L4349:             function = eval(f\"{model_location}.{module}\") | L4427:             function = eval(f\"{model_location}.{module}\") | L4562:                 source = eval(f\"{model_location}.torch\") | L4569:             function = eval(f\"source.nn.{module}\") | L4628:             exec(\nL4629:                 f\"{model_location}.torch.nn.{module}.forward = forward\",\nL4630:                 globals(),\nL4631:                 locals(),\nL4632:             ) | L4634:                 exec(\nL4635:                     f\"{model_location}.nn.{module}.forward = forward\",\nL4636:                     globals(),\nL4637:                     locals(),\nL4638:                 ) | L4642:                 exec(\nL4643:                     f\"combined_module.torch.nn.{module}.forward = forward\",\nL4644:                     globals(),\nL4645:                     locals(),\nL4646:                 ) | L4648:                     exec(\nL4649:                         f\"combined_module.nn.{module}.forward = forward\",\nL4650:                         globals(),\nL4651:                         locals(),\nL4652:                     ) | L4669:             exec(\nL4670:                 f\"{model_location}.{module} = combined_module.{module}\",\nL4671:                 globals(),\nL4672:                 locals(),\nL4673:             ) | L4683:     check_dicts = dir(eval(f\"{model_location}\")) | L4685:         item = eval(f\"{model_location}.{check}\") | L4695:                         exec(\nL4696:                             f\"{model_location}.{check}['{key}'] = combined_module.{replaced_class}\",\nL4697:                             globals(),\nL4698:                             locals(),\nL4699:                         )",
+      "evidence": "Obfusc: L1013:             _mod = __import__(model_location, fromlist=items) | L4295:         f\"   {chr(92)}{chr(92)}   /|    Num examples = {num_examples:,} | Num Epochs = {num_train_epochs:,} | Total steps = {max_steps:,}\\\\n\"\\\\\nL4296:         f\"O^O/ {chr(92)}_/ {chr(92)}    Batch size per device = {self._train_batch_size:,} | Gradient accumulation steps = {args.gradient_accumulation_steps}\\\\n\"\\\\\nL4297:         f\"{chr(92)}        /    Data Parallel GPUs = {args.world_size} | Total batch size ({self._train_batch_size} x {args.gradient_accumulation_steps} x {args.world_size}) = {total_train_batch_size: sha256:b7ea9cdbe360ad323911014caf12a9d4ec87cb36195a511080e4e484c2fb80d2\nL4298:         f' \"-____-\"     Trainable parameters = {get_model_param_count(model, trainable_only=True):,} of {get_model_param_count(model):,} ({get_model_param_count(model, trainable_only=True)/get_model_p sha256:9e832c1e7b2815aa44dc42638d821cb9df22550db88d85bd4bfb29c13f984b53 | L4296:         f\"O^O/ {chr(92)}_/ {chr(92)}    Batch size per device = {self._train_batch_size:,} | Gradient accumulation steps = {args.gradient_accumulation_steps}\\\\n\"\\\\\nL4297:         f\"{chr(92)}        /    Data Parallel GPUs = {args.world_size} | Total batch size ({self._train_batch_size} x {args.gradient_accumulation_steps} x {args.world_size}) = {total_train_batch_size: sha256:b7ea9cdbe360ad323911014caf12a9d4ec87cb36195a511080e4e484c2fb80d2\nL4298:         f' \"-____-\"     Trainable parameters = {get_model_param_count(model, trainable_only=True):,} of {get_model_param_count(model):,} ({get_model_param_count(model, trainable_only=True)/get_model_p sha256:9e832c1e7b2815aa44dc42638d821cb9df22550db88d85bd4bfb29c13f984b53 | L4295:         f\"   {chr(92)}{chr(92)}   /|    Num examples = {num_examples:,} | Num Epochs = {num_train_epochs:,} | Total steps = {max_steps:,}\\\\n\"\\\\\nL4296:         f\"O^O/ {chr(92)}_/ {chr(92)}    Batch size per device = {self._train_batch_size:,} | Gradient accumulation steps = {args.gradient_accumulation_steps}\\\\n\"\\\\\nL4297:         f\"{chr(92)}        /    Data Parallel GPUs = {args.world_size} | Total batch size ({self._train_batch_size} x {args.gradient_accumulation_steps} x {args.world_size}) = {total_train_batch_size: sha256:b7ea9cdbe360ad323911014caf12a9d4ec87cb36195a511080e4e484c2fb80d2\nExec: L612:                 if eval(_dtype) is not None: | L613:                     dtype = eval(_dtype) | L955:         _modeling_file = eval(model_location) | L1255:     f = eval(f\"{model_location}.{module}\") | L1563:         exec(f\"def raise_{j}(*args, **kwargs): print('{function}')\", globals(), locals()) | L1564:         try: exec(f\"EMPTY_LOGITS.{function} = raise_{j}\", globals(), locals()) | L2699:         exec(f\"import {parent}\", locals(), globals()) | L2830:                 dir(eval(parent)), | L2834:             exec(f\"{parent}.{child}.forward = forward\", globals(), locals()) | L2908:         module = eval(f\"modeling_file.{module}\") | L2935:             inner_class = eval(f\"modeling_file.{inner_class}\") | L3065:                 exec(f\"from timm.layers.norm_act import {norm}\") | L3073:             forward = eval(norm).forward | L3079:             exec(f\"timm.layers.norm_act.{norm}.forward = forward\") | L3096:                 exec(f\"from timm.models._efficientnet_blocks import {block}\") | L3104:             forward = eval(block).forward | L3110:             exec(f\"timm.models._efficientnet_blocks.{block}.forward = forward\") | L3389:         exec(f\"import {model_location}\", globals()) | L3392:     modeling_file = eval(model_location) | L3405:     exec(\nL3406:         \"model_logger.addFilter(HideLoggingMessage('`use_cache`'))\", globals(), locals()\nL3407:     ) | L3409:     exec(\nL3410:         \"model_logger.addFilter(HideLoggingMessage('compile_config'))\",\nL3411:         globals(),\nL3412:         locals(),\nL3413:     ) | L3564:         source = eval(f\"modeling_file.{module}\") | L3578:         source = eval(f\"modeling_file.{module}\") | L3679:         source = eval(f\"modeling_file.{module}\") | L3717:             source = eval(f\"{model_location}.{module}\") | L3788:                 source = eval(f\"{model_location}.{module}\") | L3836:             source = eval(f\"{model_location}.{module}\") | L4058:         source = eval(f\"{model_location}.{module}\") | L4069:         exec(\nL4070:             f\"{model_location}.{module}._update_causal_mask = no_update_causal_mask\",\nL4071:             globals(),\nL4072:         ) | L4135:                 source = eval(f\"{model_location}.{module}\") | L4176:                 module_cls = eval(f\"{model_location}.{module}\") | L4213:                 module_cls = eval(f\"{model_location}.{module}\") | L4280:     exec(\nL4281:         \"from transformers.trainer import (\" + \", \".join(x for x in good_items) + \")\",\nL4282:         globals(),\nL4283:     ) | L4345:     exec(inner_training_loop, globals()) | L4353:             function = eval(f\"{model_location}.{module}\") | L4431:             function = eval(f\"{model_location}.{module}\") | L4566:                 source = eval(f\"{model_location}.torch\") | L4573:             function = eval(f\"source.nn.{module}\") | L4632:             exec(\nL4633:                 f\"{model_location}.torch.nn.{module}.forward = forward\",\nL4634:                 globals(),\nL4635:                 locals(),\nL4636:             ) | L4638:                 exec(\nL4639:                     f\"{model_location}.nn.{module}.forward = forward\",\nL4640:                     globals(),\nL4641:                     locals(),\nL4642:                 ) | L4646:                 exec(\nL4647:                     f\"combined_module.torch.nn.{module}.forward = forward\",\nL4648:                     globals(),\nL4649:                     locals(),\nL4650:                 ) | L4652:                     exec(\nL4653:                         f\"combined_module.nn.{module}.forward = forward\",\nL4654:                         globals(),\nL4655:                         locals(),\nL4656:                     ) | L4673:             exec(\nL4674:                 f\"{model_location}.{module} = combined_module.{module}\",\nL4675:                 globals(),\nL4676:                 locals(),\nL4677:             ) | L4687:     check_dicts = dir(eval(f\"{model_location}\")) | L4689:         item = eval(f\"{model_location}.{check}\") | L4699:                         exec(\nL4700:                             f\"{model_location}.{check}['{key}'] = combined_module.{replaced_class}\",\nL4701:                             globals(),\nL4702:                             locals(),\nL4703:                         )",
       "evidence_hash": "ec1875fd32d00fe885e566ebda75163e46e838ca31020abb57e0991892c2bdf7"
     },
     {
@@ -1423,8 +1511,8 @@
       "file": "unsloth_zoo/mlx/loader.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L2218:             _mod = __import__(module_name, fromlist=[\"_\"])\nExec: L140:     mx.eval(model.parameters()) | L176:     mx.eval(model.parameters()) | L2022:         model.eval() | L2605:         mx.eval(model.parameters()) | L2721:         mx.eval(module.weight) | L4030:                 mx.eval(model.parameters()) | L4058:                 mx.eval(model.parameters()) | L4178:                 mx.eval(model.parameters())",
-      "evidence_hash": "9b29dade82912216c8b4808aa293b79749aa80ef1d2be35edd93bec7632810f1"
+      "evidence": "Obfusc: L2869:             _mod = __import__(module_name, fromlist=[\"_\"])\nExec: L148:     mx.eval(model.parameters()) | L180:     mx.eval(model.parameters()) | L732:                 mx.eval(model.parameters()) | L733:                 mx.eval(mx.distributed.all_sum(mx.array(1.0), stream=mx.cpu)) | L799:             mx.eval(model.parameters()) | L802:                 mx.eval(mx.distributed.all_sum(mx.array(1.0), stream=mx.cpu)) | L2673:         model.eval() | L3256:         mx.eval(model.parameters()) | L3372:         mx.eval(module.weight) | L5666:                 mx.eval(model.parameters()) | L5716:                 mx.eval(model.parameters()) | L5859:                 mx.eval(model.parameters())",
+      "evidence_hash": "7b44760032c5df6d379ccfdd0bff3d23f857f64e08210fa0fba8d2881d457634"
     },
     {
       "package": "unsloth-zoo",
@@ -1439,7 +1527,7 @@
       "file": "unsloth_zoo/saving_utils.py",
       "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
       "severity": "HIGH",
-      "evidence": "Obfusc: L3241:         module = __import__('transformers', fromlist=[model_class_name])\nExec: L3123:     exec(f\"from transformers.modeling_utils import ({', '.join(functions)})\", locals(), globals()) | L3169:     exec(save_pretrained, globals(), functions)",
+      "evidence": "Obfusc: L4015:         module = __import__('transformers', fromlist=[model_class_name])\nExec: L3897:     exec(f\"from transformers.modeling_utils import ({', '.join(functions)})\", locals(), globals()) | L3943:     exec(save_pretrained, globals(), functions)",
       "evidence_hash": "530b2383acd9fe8330aa65cd0bf86164aaacd47770e7c8d0752195bee36396ec"
     },
     {
@@ -1449,118 +1537,6 @@
       "severity": "HIGH",
       "evidence": "Obfusc: L836:         code = compile(module, \"<werkzeug routing>\", \"exec\")\nExec: L736:         exec(code, globs, locs)",
       "evidence_hash": "5c0992c90f05c772abd94d00784f157de337e1f8567f8b3aee1b15e46c96cd5d"
-    },
-    {
-      "package": "multiprocess",
-      "file": "multiprocess/tests/__init__.py",
-      "check": "Reverse shell / bind shell pattern",
-      "severity": "CRITICAL",
-      "evidence": "L3355:                     os.dup2(conn.fileno(), i) | L3387:                          \"test needs os.dup2()\") | L3405:             os.dup2(fd, newfd) | L19: import socket sha256:26a745abdc7e89da28ab943394234d8ccb415e805477c3cc1f7d4766341a4c4c",
-      "evidence_hash": "a6b9bb85e9bb6682ab0dea4f95fd9266e8802f118c76d86dd87f7ab5864872cf"
-    },
-    {
-      "package": "unsloth-zoo",
-      "file": "scripts/scan_packages.py",
-      "check": "Writes to /tmp and executes (staged dropper)",
-      "severity": "CRITICAL",
-      "evidence": "L308:     r\"/tmp/\\S+.*(?:subprocess|os\\.system|os\\.popen|Popen|chmod.*\\+x)\", | L308:     r\"/tmp/\\S+.*(?:subprocess|os\\.system|os\\.popen|Popen|chmod.*\\+x)\", sha256:78268349021e21bedcd2eaaa5b4a71b0de1d52e023ada914dfdc09515ee1aad8",
-      "evidence_hash": "590fe1c96c442fbea5eb8642650257bc0b0199e919b9bacdb11dfa767b6fe839"
-    },
-    {
-      "package": "multiprocess",
-      "file": "multiprocess/tests/__init__.py",
-      "check": "Reverse shell / bind shell pattern",
-      "severity": "CRITICAL",
-      "evidence": "L3521:                     os.dup2(conn.fileno(), i) | L3553:                          \"test needs os.dup2()\") | L3571:             os.dup2(fd, newfd) | L20: import socket sha256:07d2933301c0dbeeb6e42381687827d8dd7cfd7471986c559ca64283d5ae6e24",
-      "evidence_hash": "db1f4ca69865ec3911d7450fe11d212b817139deda21cd7a4ee32d547a8dc452"
-    },
-    {
-      "package": "fastapi",
-      "file": "fastapi/routing.py",
-      "check": "C2 polling/beaconing loop detected",
-      "severity": "CRITICAL",
-      "evidence": "L586:                                 while True: sha256:bef9ea429314fad39e063895a37dc5cfe9b04561f3d1acbb3c99abb4e92e6cfe",
-      "evidence_hash": "b15773e1bc249713156a349278ea60f7c0e3dd7d537affe929ab51089e1942bb"
-    },
-    {
-      "package": "tensorboard",
-      "file": "tensorboard/plugins/projector/tf_projector_plugin/projector_binary.js",
-      "check": "Python wheel ships large JS bundle (uncommon; manually review)",
-      "severity": "HIGH",
-      "evidence": "sha256: 53c38430766be25dc672a30846ac3b9eba86aee35eb0746785ec012647c7d9a2",
-      "evidence_hash": "2c6384e8115a6d5dacf1f84d8f724832d8dc59feb442bb98ffae0857c0ccb381"
-    },
-    {
-      "package": "fastapi",
-      "file": "fastapi/routing.py",
-      "check": "C2 polling/beaconing loop detected",
-      "severity": "CRITICAL",
-      "evidence": "L586:                                 while True: sha256:251135b5ebfdd1248916449f32262575e003ef64382501c65b7e4061d67bda45",
-      "evidence_hash": "365aef4449c8089753d9398417cd76ab762cef547d75db70d87bca9c0b550ab5"
-    },
-    {
-      "package": "fastmcp-slim",
-      "file": "fastmcp/cli/apps_dev.py",
-      "check": "Enumerates filesystem AND makes network calls",
-      "severity": "CRITICAL",
-      "evidence": "FS: L637:     history.replaceState(null, \"\", url); sha256:17068ba5bfed62c3a3007ec8bf3e0ea41ef6529b9e6112064d9afb3be9231436\nNetwork: L1304:     with httpx.Client(timeout=30.0) as client: | L1318:     with httpx.Client(timeout=30.0) as client: | L1348:     with httpx.Client(timeout=30.0) as client: | L1549:         client = httpx.AsyncClient(\nL1550:             timeout=httpx.Timeout(60.0, read=None), trust_env=False\nL1551:         ) | L1713:     async with httpx.AsyncClient(trust_env=False) as client: | L1781:                     with socket.socket(family, socket.SOCK_STREAM) as s:",
-      "evidence_hash": "e5325edfada6499540e6f0c24a0868979d275522e2b6a180aa9b5dd3280681b4"
-    },
-    {
-      "package": "huggingface-hub",
-      "file": "huggingface_hub/_sandbox.py",
-      "check": "C2 polling/beaconing loop detected",
-      "severity": "CRITICAL",
-      "evidence": "L1179:             while True: sha256:33ceddf9e42aae207e891e97808c518e92a0b27ab60e4326256717bfb25a3a38",
-      "evidence_hash": "802fd41d8bb17bf425e99d128c0351c820103a5efb74690a4086e542a71437b8"
-    },
-    {
-      "package": "huggingface-hub",
-      "file": "huggingface_hub/_sandbox.py",
-      "check": "Writes to /tmp and executes (staged dropper)",
-      "severity": "CRITICAL",
-      "evidence": "L83: d=/tmp/.sbx-server\nL84: if command -v wget >/dev/null 2>&1; then wget -q --header \"Authorization: Bearer $SBX_DL_TOKEN\" -O \"$d\" \"$SBX_SERVER_URL\"\nL85: elif command -v curl >/dev/null 2>&1; then curl -fsSL -H \"Authorization: Bearer $SBX_DL_TOKEN\" -o \"$d\" \"$SBX_SERVER_URL\"\nL86: else cp \"$SBX_SERVER_MOUNT/sbx-server\" \"$d\"; fi\nL87: chmod +x \"$d\"",
-      "evidence_hash": "6908a3fe328fa94ee22a119998d6ad07cfa1ba4efa2628acf240f4204fd76e22"
-    },
-    {
-      "package": "huggingface-hub",
-      "file": "huggingface_hub/hf_api.py",
-      "check": "C2 polling/beaconing loop detected",
-      "severity": "CRITICAL",
-      "evidence": "L4613:         while True: sha256:f764b6ca3118b23c7c0e670e77178c022a6905f825d7df6e528545fa10aae8f6",
-      "evidence_hash": "9c85d50c227285fa8dc69512999cbb082258cda4b299c7d0e0f69f5aff7accd4"
-    },
-    {
-      "package": "huggingface-hub",
-      "file": "huggingface_hub/utils/_http.py",
-      "check": "C2 polling/beaconing loop detected",
-      "severity": "CRITICAL",
-      "evidence": "L462:     while True: sha256:c75d1ee228cf7703a8c28551d649395a1f89f69a3aba69413f5bbcbd10c31958",
-      "evidence_hash": "d4d5f83fed39b87898cf776d5dad0bf1a6388a932f5fb7997d1070b50e46213e"
-    },
-    {
-      "package": "cffi",
-      "file": "cffi/_cffi_gen_src.py",
-      "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
-      "severity": "HIGH",
-      "evidence": "Obfusc: L52:     compiled = compile(source=pysrc, filename=filename, mode='exec')\nExec: L53:     exec(compiled, globs, globs)",
-      "evidence_hash": "c429e4c977a61db6b7c717b5a552fce74eda622213e49eb5467a3782fd746fb9"
-    },
-    {
-      "package": "multiprocess",
-      "file": "multiprocess/forkserver.py",
-      "check": "Reverse shell / bind shell pattern",
-      "severity": "CRITICAL",
-      "evidence": "L6: import socket sha256:6c707119169286c9a798e2c8d13a48614e481d8a503950916fd4ffb4c94d3182",
-      "evidence_hash": "50fec0f0522a8e4e636bf348b752002d7935d8455af31fb78c6f11e2eba19f6d"
-    },
-    {
-      "package": "multiprocess",
-      "file": "multiprocess/tests/__init__.py",
-      "check": "Reverse shell / bind shell pattern",
-      "severity": "CRITICAL",
-      "evidence": "L3569:                     os.dup2(conn.fileno(), i) | L3601:                          \"test needs os.dup2()\") | L3619:             os.dup2(fd, newfd) | L20: import socket sha256:c824dc0f409f242420c3fbb324790c53cb3078d2c8b07ee8f2a05694b01c2946",
-      "evidence_hash": "3878a2b430c175dbc5877a95195bfe52f9588ff73fb74e2261ed5e33087915ad"
     }
   ]
 }
diff --git a/studio/backend/core/inference/_vulkan_probe.py b/studio/backend/core/inference/_vulkan_probe.py
new file mode 100644
index 0000000000..706346daad
--- /dev/null
+++ b/studio/backend/core/inference/_vulkan_probe.py
@@ -0,0 +1,110 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Standalone free-VRAM probe for the bundled ggml Vulkan backend.
+
+Run in a short-lived subprocess (``python _vulkan_probe.py <bindir>``) so the
+Vulkan instance never lives in the long-running backend process. Loads the
+bundled ggml Vulkan backend from ``<bindir>`` and prints one
+``<idx>\\t<free_bytes>\\t<is_igpu>\\t<total_bytes>`` line per device to stdout.
+Indices are ggml's own Vulkan device ordinals, which need not match nvidia-smi
+order. ``is_igpu`` (from ggml's device type) is ``1`` for an integrated GPU
+sharing system RAM. ``total_bytes`` is the device-local heap; the reader uses
+it to reserve absolute headroom on a discrete card (parity with the CUDA/ROCm
+fit) and ignores it for an iGPU, whose "VRAM" is shared system RAM.
+
+Uses only the standard library so it stays runnable as a bare script.
+"""
+
+import ctypes
+import os
+import sys
+
+# ggml_backend_dev_type enum (ggml-backend.h): CPU=0, GPU=1, IGPU=2, ...
+_GGML_BACKEND_DEVICE_TYPE_IGPU = 2
+
+
+def _igpu_flags(base, lib, count: int) -> list[bool]:
+    """Per-device integrated-GPU flags via ggml's backend registry.
+
+    The Vulkan reg enumerates devices in the same order as
+    ``ggml_backend_vk_get_device_memory`` (each context uses ``ctx->device =
+    i``), so reg index == device ordinal. Returns all-False on any failure so
+    the reader never over-caps a discrete card.
+    """
+    flags = [False] * count
+    try:
+        lib.ggml_backend_vk_reg.restype = ctypes.c_void_p
+        lib.ggml_backend_vk_reg.argtypes = []
+        base.ggml_backend_reg_dev_count.restype = ctypes.c_size_t
+        base.ggml_backend_reg_dev_count.argtypes = [ctypes.c_void_p]
+        base.ggml_backend_reg_dev_get.restype = ctypes.c_void_p
+        base.ggml_backend_reg_dev_get.argtypes = [ctypes.c_void_p, ctypes.c_size_t]
+        base.ggml_backend_dev_type.restype = ctypes.c_int
+        base.ggml_backend_dev_type.argtypes = [ctypes.c_void_p]
+
+        reg = lib.ggml_backend_vk_reg()
+        if not reg:
+            return flags
+        dev_count = base.ggml_backend_reg_dev_count(reg)
+        for i in range(min(count, dev_count)):
+            dev = base.ggml_backend_reg_dev_get(reg, i)
+            if dev:
+                flags[i] = base.ggml_backend_dev_type(dev) == _GGML_BACKEND_DEVICE_TYPE_IGPU
+    except Exception:
+        # Best-effort: any failure degrades to "discrete" so the memory
+        # readings still get through instead of crashing the probe.
+        pass
+    return flags
+
+
+def main() -> int:
+    if len(sys.argv) < 2:
+        return 0
+    bindir = sys.argv[1]
+
+    # Hold add_dll_directory's handle for the rest of main() (the documented
+    # idiom) so bindir stays on the search path while the sibling ggml DLLs
+    # resolve below.
+    _dll_dir = None
+    if sys.platform == "win32":
+        base_name, vk_name = "ggml-base.dll", "ggml-vulkan.dll"
+        try:
+            _dll_dir = os.add_dll_directory(bindir)
+        except Exception:
+            pass
+    else:
+        base_name, vk_name = "libggml-base.so", "libggml-vulkan.so"
+
+    # RTLD_GLOBAL exposes ggml-base's symbols to ggml-vulkan on POSIX. getattr
+    # falls back to 0 where the flag doesn't exist (Windows CDLL ignores mode).
+    _rtld_global = getattr(ctypes, "RTLD_GLOBAL", 0)
+    try:
+        base = ctypes.CDLL(os.path.join(bindir, base_name), mode = _rtld_global)
+        lib = ctypes.CDLL(os.path.join(bindir, vk_name), mode = _rtld_global)
+    except OSError as e:
+        print(f"ggml-vulkan load failed: {e}", file = sys.stderr)
+        return 1
+
+    lib.ggml_backend_vk_get_device_count.restype = ctypes.c_int
+    lib.ggml_backend_vk_get_device_count.argtypes = []
+    lib.ggml_backend_vk_get_device_memory.restype = None
+    lib.ggml_backend_vk_get_device_memory.argtypes = [
+        ctypes.c_int,
+        ctypes.POINTER(ctypes.c_size_t),
+        ctypes.POINTER(ctypes.c_size_t),
+    ]
+
+    count = lib.ggml_backend_vk_get_device_count()
+    igpu = _igpu_flags(base, lib, count)
+    rows = []
+    for i in range(count):
+        free, total = ctypes.c_size_t(0), ctypes.c_size_t(0)
+        lib.ggml_backend_vk_get_device_memory(i, ctypes.byref(free), ctypes.byref(total))
+        rows.append("%d\t%d\t%d\t%d" % (i, free.value, int(igpu[i]), total.value))
+    sys.stdout.write("\n".join(rows))
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/studio/backend/core/inference/chat_template_helpers.py b/studio/backend/core/inference/chat_template_helpers.py
index dfd4c1c0bc..897db8262d 100644
--- a/studio/backend/core/inference/chat_template_helpers.py
+++ b/studio/backend/core/inference/chat_template_helpers.py
@@ -12,6 +12,43 @@ import json
 import logging
 from typing import Optional
 
+_THINK_OPEN = "<think>"
+_THINK_CLOSE = "</think>"
+
+
+def detect_think_prefill(prompt: Optional[str], special_tokens = None) -> str:
+    """Return the trailing open ``<think>`` prefill of a rendered prompt.
+
+    Reasoning templates (Qwen3.6, DeepSeek-R1-style) end the generation
+    prompt with ``<think>\\n`` so the model starts reasoning immediately.
+    Because that opening tag is part of the *prompt*, skip_prompt streaming
+    never emits it, and the frontend's ``<think>``/``</think>`` parser shows
+    the reasoning as plain text instead of a thinking block. (The GGUF path
+    is unaffected: llama-server's reasoning parser returns
+    ``reasoning_content``, which gets re-wrapped in think tags.)
+
+    Returns the exact prompt tail to re-emit at the start of the generated
+    stream (e.g. ``"<think>\\n"``), or ``""`` when the prompt does not end
+    with an open think block, including the ``enable_thinking=False`` case
+    where templates prefill an already-closed ``<think>\\n\\n</think>``.
+
+    ``special_tokens`` is the tokenizer's special-token list. If ``</think>``
+    is one, the streamer's skip_special_tokens strips the model's closing tag,
+    so re-emitting the open would leave an unclosed block that swallows the
+    answer. In that case return ``""`` and fall back to plain text.
+    """
+    if not prompt:
+        return ""
+    open_idx = prompt.rfind(_THINK_OPEN)
+    if open_idx == -1:
+        return ""
+    tail = prompt[open_idx:]
+    if _THINK_CLOSE in tail or tail.strip() != _THINK_OPEN:
+        return ""
+    if special_tokens and _THINK_CLOSE in set(special_tokens):
+        return ""
+    return tail
+
 
 logger = logging.getLogger(__name__)
 
diff --git a/studio/backend/core/inference/inference.py b/studio/backend/core/inference/inference.py
index 167706f701..7e69e05124 100644
--- a/studio/backend/core/inference/inference.py
+++ b/studio/backend/core/inference/inference.py
@@ -1178,13 +1178,22 @@ class InferenceBackend:
                 add_special_tokens = False,
                 return_tensors = "pt",
             ).to(model.device)
+            prompt_text = input_text
         else:
             # Text-only path for a vision model
             formatted_prompt = self.format_chat_prompt(messages, system_prompt)
             inputs = raw_tokenizer(formatted_prompt, return_tensors = "pt").to(model.device)
+            prompt_text = formatted_prompt
 
         # Stream with TextIteratorStreamer + background thread
         try:
+            from core.inference.chat_template_helpers import detect_think_prefill
+
+            # Re-emit an open <think> prefill swallowed by skip_prompt (see
+            # generate_stream).
+            think_prefix = detect_think_prefill(
+                prompt_text, getattr(raw_tokenizer, "all_special_tokens", None)
+            )
             from transformers import TextIteratorStreamer
             import threading
 
@@ -1233,7 +1242,11 @@ class InferenceBackend:
             thread = threading.Thread(target = generate_fn)
             thread.start()
 
-            output = ""
+            output = think_prefix
+            # Emit the prefilled <think> before the first token so the block
+            # renders during prompt prefill (which can take seconds).
+            if think_prefix:
+                yield think_prefix
             from queue import Empty
 
             generation_complete = False
@@ -1467,6 +1480,16 @@ class InferenceBackend:
 
             from transformers import TextIteratorStreamer
             import threading
+            from core.inference.chat_template_helpers import detect_think_prefill
+
+            # skip_prompt swallows an open <think> prefilled by the template;
+            # re-emit it so the frontend can render the thinking block.
+            # gpt-oss emits its own tags via HarmonyTextStreamer.
+            think_prefix = (
+                ""
+                if self._is_gpt_oss_model()
+                else detect_think_prefill(prompt, getattr(tokenizer, "all_special_tokens", None))
+            )
 
             # gpt-oss models: HarmonyTextStreamer parses the multi-channel
             # harmony protocol into <think> tags
@@ -1550,7 +1573,11 @@ class InferenceBackend:
             thread = threading.Thread(target = generate_fn)
             thread.start()
 
-            output = ""
+            output = think_prefix
+            # Emit the prefilled <think> before the first token so the block
+            # renders during prompt prefill (which can take seconds).
+            if think_prefix:
+                yield think_prefix
             from queue import Empty
 
             generation_complete = False
diff --git a/studio/backend/core/inference/llama_admission.py b/studio/backend/core/inference/llama_admission.py
new file mode 100644
index 0000000000..b6a939c87b
--- /dev/null
+++ b/studio/backend/core/inference/llama_admission.py
@@ -0,0 +1,368 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+"""Admission control for local llama-server generation requests.
+
+The helpers in this module deliberately know nothing about FastAPI, SSE, or the
+OpenAI-compatible route shape. They only coordinate how many upstream generation
+requests may be active for one llama-server backend and provide a cancellable
+FIFO queue for excess requests.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import os
+import threading
+from collections import deque
+from dataclasses import dataclass
+from typing import Deque, Optional
+
+
+ADMISSION_CONTROL_ENV = "UNSLOTH_OPENAI_COMPAT_ADMISSION_CONTROL"
+ADMISSION_QUEUE_TIMEOUT_ENV = "UNSLOTH_OPENAI_COMPAT_ADMISSION_QUEUE_TIMEOUT"
+ADMISSION_KEEPALIVE_INTERVAL_ENV = "UNSLOTH_OPENAI_COMPAT_ADMISSION_KEEPALIVE_INTERVAL"
+ADMISSION_MAX_QUEUE_ENV = "UNSLOTH_OPENAI_COMPAT_ADMISSION_MAX_QUEUE"
+
+DEFAULT_ADMISSION_ENABLED = True
+DEFAULT_ADMISSION_QUEUE_TIMEOUT_S = None
+DEFAULT_ADMISSION_KEEPALIVE_INTERVAL_S = 5.0
+DEFAULT_ADMISSION_MAX_QUEUE = 64
+
+
+@dataclass(frozen = True)
+class LlamaAdmissionConfig:
+    enabled: bool = DEFAULT_ADMISSION_ENABLED
+    queue_timeout_s: Optional[float] = DEFAULT_ADMISSION_QUEUE_TIMEOUT_S
+    keepalive_interval_s: float = DEFAULT_ADMISSION_KEEPALIVE_INTERVAL_S
+    max_queue: Optional[int] = DEFAULT_ADMISSION_MAX_QUEUE
+
+
+@dataclass(frozen = True)
+class LlamaAdmissionSnapshot:
+    key: str
+    capacity: int
+    active: int
+    queued: int
+
+
+class LlamaAdmissionError(Exception):
+    def __init__(
+        self,
+        message: str,
+        *,
+        snapshot: Optional[LlamaAdmissionSnapshot] = None,
+    ):
+        super().__init__(message)
+        self.snapshot = snapshot
+
+
+class LlamaAdmissionQueueFull(LlamaAdmissionError):
+    pass
+
+
+class LlamaAdmissionTimeout(LlamaAdmissionError):
+    pass
+
+
+class LlamaAdmissionCancelled(LlamaAdmissionError):
+    pass
+
+
+def _bool_env(name: str, default: bool) -> bool:
+    value = os.environ.get(name)
+    if value is None or not value.strip():
+        return default
+    value = value.strip().lower()
+    if value in {"1", "true", "yes", "on"}:
+        return True
+    if value in {"0", "false", "no", "off"}:
+        return False
+    return default
+
+
+def _optional_positive_float_env(name: str, default: Optional[float]) -> Optional[float]:
+    value = os.environ.get(name)
+    if value is None or not value.strip():
+        return default
+    try:
+        parsed = float(value.strip())
+    except ValueError:
+        return default
+    return parsed if parsed > 0 else None
+
+
+def _positive_float_env(name: str, default: float) -> float:
+    value = os.environ.get(name)
+    if value is None or not value.strip():
+        return default
+    try:
+        parsed = float(value.strip())
+    except ValueError:
+        return default
+    return parsed if parsed > 0 else default
+
+
+def _optional_positive_int_env(name: str, default: Optional[int]) -> Optional[int]:
+    value = os.environ.get(name)
+    if value is None or not value.strip():
+        return default
+    try:
+        parsed = int(value.strip())
+    except ValueError:
+        return default
+    return parsed if parsed > 0 else None
+
+
+def llama_admission_config_from_env() -> LlamaAdmissionConfig:
+    return LlamaAdmissionConfig(
+        enabled = _bool_env(ADMISSION_CONTROL_ENV, DEFAULT_ADMISSION_ENABLED),
+        queue_timeout_s = _optional_positive_float_env(
+            ADMISSION_QUEUE_TIMEOUT_ENV,
+            DEFAULT_ADMISSION_QUEUE_TIMEOUT_S,
+        ),
+        keepalive_interval_s = _positive_float_env(
+            ADMISSION_KEEPALIVE_INTERVAL_ENV,
+            DEFAULT_ADMISSION_KEEPALIVE_INTERVAL_S,
+        ),
+        max_queue = _optional_positive_int_env(
+            ADMISSION_MAX_QUEUE_ENV,
+            DEFAULT_ADMISSION_MAX_QUEUE,
+        ),
+    )
+
+
+@dataclass
+class _Waiter:
+    loop: asyncio.AbstractEventLoop
+    future: asyncio.Future
+    cancelled: bool = False
+    granted_lease: Optional["LlamaAdmissionLease"] = None
+
+
+class LlamaAdmissionLease:
+    def __init__(self, queue: Optional["LlamaAdmissionQueue"]):
+        self._queue = queue
+        self._released = False
+        self._release_lock = threading.Lock()
+
+    def release(self) -> None:
+        queue = None
+        with self._release_lock:
+            if self._released:
+                return
+            self._released = True
+            queue = self._queue
+        if queue is not None:
+            queue.release()
+
+    async def __aenter__(self) -> "LlamaAdmissionLease":
+        return self
+
+    async def __aexit__(self, *_args) -> None:
+        self.release()
+
+
+class LlamaAdmissionReservation:
+    def __init__(
+        self,
+        *,
+        queue: Optional["LlamaAdmissionQueue"],
+        lease: Optional[LlamaAdmissionLease] = None,
+        waiter: Optional[_Waiter] = None,
+        snapshot: Optional[LlamaAdmissionSnapshot] = None,
+    ):
+        self._queue = queue
+        self._lease = lease
+        self._waiter = waiter
+        self.snapshot = snapshot
+
+    @property
+    def is_cancelled(self) -> bool:
+        return self._lease is None and self._waiter is None
+
+    def lease_nowait(self) -> Optional[LlamaAdmissionLease]:
+        if self._lease is not None:
+            return self._lease
+        if self._waiter is None or not self._waiter.future.done():
+            return None
+        if self._waiter.future.cancelled():
+            self._waiter.cancelled = True
+            self._waiter = None
+            return None
+        self._lease = self._waiter.future.result()
+        self._waiter = None
+        return self._lease
+
+    async def wait(self, timeout_s: float) -> Optional[LlamaAdmissionLease]:
+        lease = self.lease_nowait()
+        if lease is not None:
+            return lease
+        if self._waiter is None:
+            return None
+        waiter = self._waiter
+        try:
+            await asyncio.wait_for(asyncio.shield(waiter.future), timeout = timeout_s)
+        except asyncio.CancelledError:
+            if waiter.future.cancelled():
+                waiter.cancelled = True
+                if self._waiter is waiter:
+                    self._waiter = None
+                return None
+            raise
+        return self.lease_nowait()
+
+    def cancel(self) -> None:
+        lease = self.lease_nowait()
+        if lease is not None:
+            lease.release()
+            self._lease = None
+            return
+        if self._queue is not None and self._waiter is not None:
+            self._queue.cancel(self._waiter)
+        self._waiter = None
+
+    def snapshot_now(self) -> Optional[LlamaAdmissionSnapshot]:
+        if self._queue is None:
+            return self.snapshot
+        return self._queue.snapshot()
+
+
+class LlamaAdmissionQueue:
+    def __init__(self, key: str):
+        self.key = key
+        self._lock = threading.Lock()
+        self._active = 0
+        self._capacity = 1
+        self._waiters: Deque[_Waiter] = deque()
+
+    def reserve(self, *, capacity: int, config: LlamaAdmissionConfig) -> LlamaAdmissionReservation:
+        capacity = max(1, int(capacity or 1))
+        if not config.enabled:
+            return LlamaAdmissionReservation(
+                queue = None,
+                lease = LlamaAdmissionLease(None),
+                snapshot = LlamaAdmissionSnapshot(self.key, capacity, 0, 0),
+            )
+
+        loop = asyncio.get_running_loop()
+        with self._lock:
+            self._capacity = capacity
+            self._prune_waiters_locked()
+            self._grant_waiters_locked()
+            if self._active < self._capacity and not self._waiters:
+                self._active += 1
+                return LlamaAdmissionReservation(
+                    queue = self,
+                    lease = LlamaAdmissionLease(self),
+                    snapshot = self._snapshot_locked(),
+                )
+            if config.max_queue is not None and len(self._waiters) >= config.max_queue:
+                raise LlamaAdmissionQueueFull(
+                    "llama-server generation queue is full",
+                    snapshot = self._snapshot_locked(),
+                )
+            waiter = _Waiter(
+                loop = loop,
+                future = loop.create_future(),
+            )
+            self._waiters.append(waiter)
+            return LlamaAdmissionReservation(
+                queue = self,
+                waiter = waiter,
+                snapshot = self._snapshot_locked(),
+            )
+
+    def release(self) -> None:
+        with self._lock:
+            if self._active > 0:
+                self._active -= 1
+            self._grant_waiters_locked()
+
+    def cancel(self, waiter: _Waiter) -> None:
+        lease_to_release = None
+        with self._lock:
+            waiter.cancelled = True
+            try:
+                self._waiters.remove(waiter)
+            except ValueError:
+                pass
+            if waiter.granted_lease is not None:
+                lease_to_release = waiter.granted_lease
+                waiter.granted_lease = None
+            if not waiter.future.done():
+                waiter.loop.call_soon_threadsafe(waiter.future.cancel)
+        if lease_to_release is not None:
+            lease_to_release.release()
+
+    def snapshot(self) -> LlamaAdmissionSnapshot:
+        with self._lock:
+            self._prune_waiters_locked()
+            return self._snapshot_locked()
+
+    def is_idle(self) -> bool:
+        with self._lock:
+            self._prune_waiters_locked()
+            return self._active == 0 and not self._waiters
+
+    def _grant_waiters_locked(self) -> None:
+        self._prune_waiters_locked()
+        while self._waiters and self._active < self._capacity:
+            waiter = self._waiters.popleft()
+            if waiter.cancelled or waiter.future.done():
+                continue
+            self._active += 1
+            lease = LlamaAdmissionLease(self)
+            waiter.granted_lease = lease
+            waiter.loop.call_soon_threadsafe(self._deliver_lease, waiter, lease)
+
+    def _deliver_lease(self, waiter: _Waiter, lease: LlamaAdmissionLease) -> None:
+        if waiter.cancelled or waiter.future.done():
+            waiter.granted_lease = None
+            if not waiter.future.done():
+                waiter.future.cancel()
+            lease.release()
+            return
+        try:
+            waiter.future.set_result(lease)
+            waiter.granted_lease = None
+        except asyncio.InvalidStateError:
+            waiter.granted_lease = None
+            lease.release()
+
+    def _prune_waiters_locked(self) -> None:
+        self._waiters = deque(
+            waiter for waiter in self._waiters if not waiter.cancelled and not waiter.future.done()
+        )
+
+    def _snapshot_locked(self) -> LlamaAdmissionSnapshot:
+        return LlamaAdmissionSnapshot(
+            key = self.key,
+            capacity = self._capacity,
+            active = self._active,
+            queued = len(self._waiters),
+        )
+
+
+_QUEUES_LOCK = threading.Lock()
+_QUEUES: dict[str, LlamaAdmissionQueue] = {}
+
+
+def get_llama_admission_queue(key: str) -> LlamaAdmissionQueue:
+    with _QUEUES_LOCK:
+        queue = _QUEUES.get(key)
+        if queue is None:
+            queue = LlamaAdmissionQueue(key)
+            _QUEUES[key] = queue
+            # base_url carries a fresh ephemeral port on every model load, so
+            # each load registers a new key. Drop the now-idle queues from prior
+            # loads so the registry can't grow without bound on a long-running
+            # server. Queues with in-flight requests are kept until they drain.
+            for stale_key in [k for k in _QUEUES if k != key and _QUEUES[k].is_idle()]:
+                del _QUEUES[stale_key]
+        return queue
+
+
+def reset_llama_admission_queues() -> None:
+    with _QUEUES_LOCK:
+        _QUEUES.clear()
diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 757467a008..b06c6eb5cb 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -67,7 +67,6 @@ from core.tool_healing import (
     _strip_bracket_tag_calls,
     apply_tool_strip_patterns,
     strip_outside_think,
-    strip_tool_call_markup,
 )
 from utils.native_path_leases import child_env_without_native_path_secret
 from utils.hf_xet_fallback import hf_hub_download_with_xet_fallback
@@ -101,6 +100,10 @@ class LlamaServerNotFoundError(RuntimeError):
     Subclasses RuntimeError so existing handlers still catch it."""
 
 
+class _LlamaStreamCancelled(Exception):
+    """Internal signal for an expected client/request cancellation."""
+
+
 # Shared so the from_identifier preflight and the load-time raise stay in sync.
 LLAMA_SERVER_NOT_FOUND_DETAIL = (
     "This is a GGUF model, but the llama.cpp runtime (llama-server) is not "
@@ -378,6 +381,19 @@ def _probe_dns_dead(host: str = "huggingface.co", timeout: float = 2.0) -> bool:
     return True if result[0] is None else result[0]
 
 
+def _hf_env_offline() -> bool:
+    """True when an HF offline env var is set to any truthy value (1/true/yes/on).
+
+    Mirrors utils.models.model_config._env_offline so a user-set HF_HUB_OFFLINE=true
+    (not just "1") still routes through the local-cache reuse path below.
+    """
+    try:
+        from utils.models.model_config import _env_offline
+        return _env_offline()
+    except Exception:
+        return os.environ.get("HF_HUB_OFFLINE", "").strip().lower() in {"1", "true", "yes", "on"}
+
+
 @contextlib.contextmanager
 def _hf_offline_if_dns_dead():
     """Set HF_HUB_OFFLINE for this block only when DNS to huggingface.co fails;
@@ -838,6 +854,112 @@ def _gguf_snapshot_files(snapshot: Path) -> list[str]:
     ]
 
 
+def _cached_hf_snapshot_file(
+    repo_id: str,
+    filename: str,
+    *,
+    expected_size: Optional[int] = None,
+) -> Optional[str]:
+    """Return a cached snapshot file even when HF's current-ref probe misses it."""
+    if not filename:
+        return None
+    parts = [part for part in filename.replace("\\", "/").split("/") if part]
+    if not parts or any(part in (".", "..") for part in parts):
+        return None
+    try:
+        from utils.models.model_config import _iter_hf_cache_snapshots
+        for snap in _iter_hf_cache_snapshots(repo_id):
+            candidate = snap.joinpath(*parts)
+            if not candidate.is_file():
+                continue
+            if expected_size:
+                try:
+                    if candidate.stat().st_size < expected_size:
+                        continue
+                except OSError:
+                    continue
+            return str(candidate)
+    except Exception as e:
+        logger.debug("Snapshot cache lookup failed for %s/%s: %s", repo_id, filename, e)
+    return None
+
+
+def _snapshot_has_all_shards(
+    main_path: str, main_filename: str, shards: Iterable[str], expected_sizes: dict[str, int]
+) -> bool:
+    """True when every shard sits beside ``main_path`` in the same cache snapshot.
+
+    llama.cpp loads a split GGUF by resolving its siblings from the main shard's
+    directory, so a cached main shard is only safe to reuse when the rest of the
+    set is co-located; otherwise the caller must fetch the whole set together.
+    """
+    root = Path(main_path)
+    for _ in [part for part in main_filename.replace("\\", "/").split("/") if part]:
+        root = root.parent
+    for shard in shards:
+        parts = [part for part in shard.replace("\\", "/").split("/") if part]
+        if not parts or any(part in (".", "..") for part in parts):
+            return False
+        sibling = root.joinpath(*parts)
+        try:
+            if not sibling.is_file():
+                return False
+            expected = expected_sizes.get(shard)
+            if expected and sibling.stat().st_size < expected:
+                return False
+        except OSError:
+            return False
+    return True
+
+
+def _resolve_repo_id_casing(hf_repo: str) -> str:
+    """Map a requested repo id to its cached canonical casing, or return it unchanged.
+
+    A case-variant request (for example a lowercased id) resolves to the
+    canonical-cased cache directory so the main GGUF and its companions
+    (mmproj / MTP drafter) all read the same cache entry. Returns ``hf_repo``
+    unchanged when resolution is unavailable or errors.
+    """
+    try:
+        from utils.paths import resolve_cached_repo_id_case
+        return resolve_cached_repo_id_case(hf_repo)
+    except Exception:
+        return hf_repo
+
+
+def _cached_colocated_split_main(
+    repo_id: str, main_filename: str, shards: Iterable[str], expected_sizes: dict[str, int]
+) -> Optional[str]:
+    """Main-shard path from a cache snapshot that also holds every sibling shard.
+
+    A newer snapshot may hold only the first shard while an older snapshot has the
+    complete split set. ``_cached_hf_snapshot_file`` would return that newer partial
+    main and the co-location check would then force a refetch, so scan snapshots for
+    one where the whole set is present and return that main path instead. None when
+    no snapshot holds the full set.
+    """
+    main_parts = [part for part in main_filename.replace("\\", "/").split("/") if part]
+    if not main_parts or any(part in (".", "..") for part in main_parts):
+        return None
+    try:
+        from utils.models.model_config import _iter_hf_cache_snapshots
+        for snap in _iter_hf_cache_snapshots(repo_id):
+            main_path = snap.joinpath(*main_parts)
+            if not main_path.is_file():
+                continue
+            expected_main = expected_sizes.get(main_filename)
+            try:
+                if expected_main and main_path.stat().st_size < expected_main:
+                    continue
+            except OSError:
+                continue
+            if _snapshot_has_all_shards(str(main_path), main_filename, shards, expected_sizes):
+                return str(main_path)
+    except Exception as e:
+        logger.debug("Co-located split snapshot lookup failed for %s: %s", repo_id, e)
+    return None
+
+
 def _gguf_extra_shards(files: Iterable[str], first_shard: str) -> list[str]:
     m = _SHARD_FULL_RE.match(first_shard)
     if not m:
@@ -1318,6 +1440,50 @@ def _backfill_usage_from_timings(usage, timings):
     return out
 
 
+def _vulkan_lib_filename() -> str:
+    return "ggml-vulkan.dll" if sys.platform == "win32" else "libggml-vulkan.so"
+
+
+# Host RAM to leave free on an integrated GPU, matching llama.cpp's own --fit
+# margin (default 1024 MiB per device). ggml reports an iGPU's "VRAM" as shared
+# system RAM, so hold back the same margin rather than inventing a larger one.
+_IGPU_HOST_RESERVE_MIB = 1024
+
+
+def _apply_igpu_host_reserve_mib(free_mib: int, is_igpu: bool) -> int:
+    """Reserve host headroom on an integrated (shared-memory) Vulkan GPU.
+
+    An iGPU's reported free "VRAM" is really free system RAM, so sizing
+    context/offload against all of it would push the host into swap or the OOM
+    killer. Leave the same margin llama.cpp's --fit uses. ``is_igpu`` comes from
+    ggml's device type, so a discrete card is never touched; only ever reduces.
+    """
+    if not is_igpu:
+        return free_mib
+    return max(0, free_mib - _IGPU_HOST_RESERVE_MIB)
+
+
+def _llama_lib_dir(binary: str) -> Path:
+    # The installer exposes llama-server as a top-level entrypoint into build/bin/,
+    # where the ggml backend libs live, so callers looking for sibling libs (Vulkan
+    # detection, LD_LIBRARY_PATH, probe bindir) need the real dir. It is normally a
+    # symlink (resolve() reaches build/bin), but create_exec_entrypoint falls back to
+    # a shell wrapper (exec "$(dirname "$0")/build/bin/llama-server" "$@") when it
+    # cannot symlink, and resolve() stops at the wrapper file. Follow the wrapper's
+    # exec target too, so a wrapper-based install still finds build/bin.
+    resolved = Path(binary).resolve()
+    try:
+        with open(resolved, "rb") as _f:
+            _head = _f.read(256)
+        if _head.startswith(b"#!"):
+            _m = re.search(r'exec "\$\(dirname "\$0"\)/([^"]+)"', _head.decode("utf-8", "ignore"))
+            if _m:
+                return (resolved.parent / _m.group(1)).resolve().parent
+    except OSError:
+        pass
+    return resolved.parent
+
+
 def _is_external_link(path: Path) -> bool:
     """True when ``path`` is a --with-llama-cpp-dir local link: a POSIX symlink
     or a Windows directory junction / reparse point. Such a link resolves into
@@ -1375,6 +1541,7 @@ class LlamaCppBackend:
         self._context_length: Optional[int] = None
         self._effective_context_length: Optional[int] = None
         self._max_context_length: Optional[int] = None
+        self._effective_parallel_slots: int = 1
         self._chat_template: Optional[str] = None
         self._chat_template_override: Optional[str] = None
         self._supports_reasoning: bool = False
@@ -1560,6 +1727,15 @@ class LlamaCppBackend:
         """Return the effective context length the server is running at."""
         return self._effective_context_length or self._context_length
 
+    @property
+    def effective_parallel_slots(self) -> int:
+        """Return the serving-slot count the active llama-server actually uses."""
+        try:
+            slots = int(getattr(self, "_effective_parallel_slots", 1))
+        except (TypeError, ValueError):
+            slots = 1
+        return max(1, slots)
+
     @property
     def max_context_length(self) -> Optional[int]:
         """Return the largest context that fits on this hardware at load time.
@@ -1576,6 +1752,16 @@ class LlamaCppBackend:
         """Return the model's native context length from GGUF metadata."""
         return self._context_length
 
+    def _commit_effective_parallel_slots(self, n_parallel: int) -> None:
+        try:
+            slots = int(n_parallel)
+        except (TypeError, ValueError):
+            slots = 1
+        self._effective_parallel_slots = max(1, slots)
+
+    def _reset_effective_parallel_slots(self) -> None:
+        self._effective_parallel_slots = 1
+
     @staticmethod
     def _read_rss_bytes(pid: int) -> Optional[int]:
         """Resident set size of ``pid`` in bytes, from /proc/<pid>/status (Linux).
@@ -1781,6 +1967,13 @@ class LlamaCppBackend:
             return False
         return self._supports_tools
 
+    @property
+    def supports_tool_passthrough(self) -> bool:
+        # supports_tools is forced off for DiffusionGemma (its agentic loop drops the
+        # per-step canvas frames), but client passthrough skips that loop, so it uses
+        # the real _supports_tools.
+        return self._supports_tools
+
     @property
     def cache_type_kv(self) -> Optional[str]:
         return self._cache_type_kv
@@ -2153,6 +2346,30 @@ class LlamaCppBackend:
 
         return total
 
+    @staticmethod
+    def _is_vulkan_backend(binary: Optional[str] = None) -> bool:
+        """True if the installed llama.cpp build is Vulkan-only.
+
+        The official prebuilts are single-backend, so the Vulkan ggml lib next
+        to llama-server identifies a Vulkan build. Keeps the free-memory probe
+        and GPU pin in ggml's Vulkan device-index space. For a custom
+        multi-backend build with a CUDA or HIP ggml lib alongside Vulkan, defer
+        to that backend (torch-usable, better-understood probe/pin).
+        """
+        binary = binary or LlamaCppBackend._find_llama_server_binary()
+        if not binary:
+            return False
+        lib_dir = _llama_lib_dir(binary)
+        if not (lib_dir / _vulkan_lib_filename()).is_file():
+            return False
+        for _backend in ("cuda", "hip"):
+            sibling = (
+                f"ggml-{_backend}.dll" if sys.platform == "win32" else f"libggml-{_backend}.so"
+            )
+            if (lib_dir / sibling).is_file():
+                return False
+        return True
+
     @staticmethod
     def _resolve_visible_physical_ids() -> Optional[list[int]]:
         """Physical GPU ids behind the active visibility mask (HIP/ROCR/CUDA on
@@ -2315,11 +2532,42 @@ class LlamaCppBackend:
         return True
 
     @staticmethod
-    def _get_gpu_free_memory() -> list[tuple[int, int]]:
+    def _visible_devices_mask(env_name: str) -> Optional[set[int]]:
+        """Physical indices a ``*_VISIBLE_DEVICES`` mask permits, or None if unset.
+
+        ``if x.strip()`` filters trailing-comma masks ("0,1,"); an empty mask
+        ("") yields an empty set (all devices hidden), distinct from an unset
+        var (None, no mask). Used by the nvidia-smi probe.
+        """
+        raw = os.environ.get(env_name)
+        if raw is None:
+            return None
+        try:
+            return set(int(x.strip()) for x in raw.split(",") if x.strip())
+        except ValueError:
+            return None
+
+    @staticmethod
+    def _vulkan_pin_args(gpu_indices: Optional[Iterable[int]]) -> list[str]:
+        """``--device Vulkan<i>,...`` to pin a Vulkan launch to selected GPUs.
+
+        The indices are ggml's compact Vulkan ordinals (as _get_gpu_free_memory
+        reports and the registry names ``Vulkan<i>``). Pin by that name, NOT via
+        GGML_VK_VISIBLE_DEVICES: ggml parses that env var in the raw
+        vkEnumeratePhysicalDevices space (before dropping CPU/llvmpipe devices
+        and deduplicating ICDs), so a compact ordinal there could select a
+        different physical device or the CPU rasterizer.
+        """
+        if not gpu_indices:
+            return []
+        return ["--device", ",".join(f"Vulkan{i}" for i in gpu_indices)]
+
+    @staticmethod
+    def _get_gpu_free_memory(binary: Optional[str] = None) -> list[tuple[int, int]]:
         """Query free memory per GPU. Returns ``(gpu_index, free_mib)`` sorted by
         index; empty if no supported GPU is reachable. Thin wrapper over
         ``_get_gpu_memory`` for callers that only need free VRAM."""
-        return [(idx, free) for idx, free, _total in LlamaCppBackend._get_gpu_memory()]
+        return [(idx, free) for idx, free, _total in LlamaCppBackend._get_gpu_memory(binary)]
 
     @staticmethod
     def _apple_metal_memory_budget_bytes() -> int:
@@ -2350,7 +2598,7 @@ class LlamaCppBackend:
         return int(rec_bytes * _APPLE_UNIFIED_MEMORY_FRACTION)
 
     @staticmethod
-    def _get_gpu_memory() -> list[tuple[int, int, int]]:
+    def _get_gpu_memory(binary: Optional[str] = None) -> list[tuple[int, int, int]]:
         """Query free AND total memory per GPU.
 
         Order:
@@ -2362,9 +2610,18 @@ class LlamaCppBackend:
              probe returned [] on AMD) and NVIDIA hosts missing
              ``nvidia-smi`` from PATH.
 
+        On a Vulkan build the ggml Vulkan probe is authoritative, so the indices
+        are ggml's compact Vulkan ordinals (the space the pin selects via
+        ``--device Vulkan<i>``). It reports ``total`` for discrete cards and 0
+        for an iGPU (shared RAM) so the fit falls back to free*frac there.
+        Otherwise nvidia-smi / torch cover NVIDIA + AMD ROCm.
+
         Returns (gpu_index, free_mib, total_mib) sorted by index; empty if no
-        supported GPU is reachable. ``total`` lets the fit reserve absolute headroom.
+        supported GPU is reachable.
         """
+        binary = binary or LlamaCppBackend._find_llama_server_binary()
+        if LlamaCppBackend._is_vulkan_backend(binary):
+            return LlamaCppBackend._get_gpu_free_memory_vulkan(binary)
         # ── NVIDIA via nvidia-smi ────────────────────────────────────
         try:
             result = subprocess.run(
@@ -2380,16 +2637,7 @@ class LlamaCppBackend:
                 **_windows_hidden_subprocess_kwargs(),
             )
             if result.returncode == 0:
-                allowed: Optional[set[int]] = None
-                cvd = os.environ.get("CUDA_VISIBLE_DEVICES")
-                if cvd is not None:
-                    try:
-                        # `if x.strip()` filters trailing-comma masks ("0,1,").
-                        # Empty mask (CVD="") yields an empty set -> all GPUs
-                        # filtered out, per codebase convention.
-                        allowed = set(int(x.strip()) for x in cvd.split(",") if x.strip())
-                    except ValueError:
-                        pass
+                allowed = LlamaCppBackend._visible_devices_mask("CUDA_VISIBLE_DEVICES")
                 gpus: list[tuple[int, int, int]] = []
                 for line in result.stdout.strip().splitlines():
                     parts = [p.strip() for p in line.split(",")]
@@ -2454,6 +2702,91 @@ class LlamaCppBackend:
             logger.debug(f"torch GPU probe failed: {e}")
             return []
 
+    @staticmethod
+    def _get_gpu_free_memory_vulkan(binary: Optional[str] = None) -> list[tuple[int, int, int]]:
+        """Query free (and total) VRAM per device via the bundled ggml Vulkan backend.
+
+        Loads ``libggml-vulkan`` in a short-lived subprocess (no Vulkan instance
+        in this process) and returns (device_index, free_mib, total_mib) sorted
+        by index. The index is ggml's compact Vulkan ordinal -- the one the
+        registry names ``Vulkan<index>`` and load_model pins with ``--device``,
+        NOT the raw ``GGML_VK_VISIBLE_DEVICES`` space. A user-set
+        ``GGML_VK_VISIBLE_DEVICES`` is honored by ggml (passed through), so the
+        list already reflects it. iGPUs leave a host-RAM margin (see
+        ``_apply_igpu_host_reserve_mib``) and report total 0; discrete cards pass
+        their real total through. [] when no Vulkan build or device is reachable.
+        """
+        binary = binary or LlamaCppBackend._find_llama_server_binary()
+        if not binary:
+            return []
+        binary_dir = _llama_lib_dir(binary)
+        if not (binary_dir / _vulkan_lib_filename()).is_file():
+            return []
+
+        env = child_env_without_native_path_secret()
+        # Pass any inherited GGML_VK_VISIBLE_DEVICES through to ggml unchanged so
+        # the probe enumerates the same device list the launch will, named
+        # Vulkan0..N in the compact order reported here and pinned by that name
+        # via --device -- probe, mask, and pin stay in one index space. Do NOT
+        # filter the mask in Python: ggml parses the env var in raw
+        # vkEnumeratePhysicalDevices space while this probe reports the compact
+        # post-filter ordinal, so a Python filter would compare mismatched spaces.
+        if sys.platform != "win32":
+            # Let the loader resolve sibling ggml libs next to the binary.
+            existing_ld = env.get("LD_LIBRARY_PATH", "")
+            env["LD_LIBRARY_PATH"] = (
+                f"{binary_dir}:{existing_ld}" if existing_ld else str(binary_dir)
+            )
+        probe_script = Path(__file__).with_name("_vulkan_probe.py")
+        try:
+            result = subprocess.run(
+                [sys.executable, str(probe_script), str(binary_dir)],
+                capture_output = True,
+                text = True,
+                timeout = 15,
+                env = env,
+                **_windows_hidden_subprocess_kwargs(),
+            )
+            if result.returncode != 0:
+                logger.debug(
+                    f"vulkan GPU probe exited {result.returncode}: {result.stderr.strip()}"
+                )
+                return []
+        except Exception as e:
+            logger.debug(f"vulkan GPU probe failed: {e}")
+            return []
+
+        gpus: list[tuple[int, int, int]] = []
+        for line in result.stdout.strip().splitlines():
+            parts = line.split("\t")
+            if len(parts) != 4:
+                continue
+            try:
+                idx = int(parts[0])
+                free_mib = int(parts[1]) // (1024 * 1024)
+                is_igpu = parts[2] == "1"
+                # iGPU "total" is shared RAM, not a VRAM budget -> keep 0 so the
+                # fit stays on free*frac (the host reserve below is its
+                # headroom); a discrete card passes its real total through.
+                total_mib = 0 if is_igpu else int(parts[3]) // (1024 * 1024)
+            except ValueError:
+                continue
+            capped = _apply_igpu_host_reserve_mib(free_mib, is_igpu)
+            if capped < free_mib:
+                logger.info(
+                    f"Vulkan device VK{idx} is an integrated GPU sharing system "
+                    f"RAM; reserving {free_mib - capped}MiB host headroom "
+                    f"({free_mib}->{capped}MiB usable)"
+                )
+            gpus.append((idx, capped, total_mib))
+        gpus.sort(key = lambda g: g[0])
+        if gpus:
+            logger.info(
+                "Vulkan GPU memory detected: "
+                + ", ".join(f"VK{idx}={free}MiB" for idx, free, _total in gpus)
+            )
+        return gpus
+
     @staticmethod
     def _available_system_memory_mib() -> Optional[int]:
         """Available system RAM in MiB (psutil, then /proc/meminfo), or None if
@@ -2682,7 +3015,8 @@ class LlamaCppBackend:
     def _llama_server_env_for_binary(binary: str) -> dict[str, str]:
         """Build a subprocess env that lets llama-server resolve native libs."""
         env = child_env_without_native_path_secret()
-        binary_dir = str(Path(binary).parent)
+        # _llama_lib_dir resolves the llama-server symlink to the real build/bin.
+        binary_dir = str(_llama_lib_dir(binary))
 
         if sys.platform == "win32":
             # Ordering: see _build_windows_path_dirs. #5106.
@@ -3846,7 +4180,11 @@ class LlamaCppBackend:
         # Auto-size (0): the visual server probes the largest context that fits this GPU's VRAM
         # (capped at the training context). An explicit in-range n_ctx overrides it.
         maxtok = n_ctx if (n_ctx and 0 < n_ctx <= 65536) else 0
-        gpu = os.environ.get("DG_GPU", "0")
+        # No visible CUDA GPU: a genuine CPU host, or a GPU host masked with
+        # CUDA_VISIBLE_DEVICES="" to force CPU serving. Keep the visual-server child
+        # CPU-masked (empty --gpu) so the shim does not re-expose GPU 0 via its default.
+        cpu_only = self._effective_gpu_count() == 0
+        gpu = "" if cpu_only else os.environ.get("DG_GPU", "0")
 
         cmd = list(shim_cmd) + [
             "--gguf",
@@ -3866,6 +4204,12 @@ class LlamaCppBackend:
         # refuses to load unless UNSLOTH_IS_PRESENT is set (normally by `import
         # unsloth`). The shim never imports unsloth, so set it here as unsloth does.
         env["UNSLOTH_IS_PRESENT"] = "1"
+        # The shim's `import unsloth_zoo` aborts in get_device_type() ("needs a GPU")
+        # when no accelerator is visible, even though it only drives the CPU
+        # visual-server binary and does no torch GPU work. Allow the CPU device so the
+        # runner starts; the visual server still runs on the CPU llama.cpp build.
+        if cpu_only:
+            env.setdefault("UNSLOTH_ALLOW_CPU", "1")
         env["DG_VISUAL_BIN"] = visual_bin
         env["DG_GPU"] = gpu
         # The file-override shim imports its sibling visual_engine; put its dir on PYTHONPATH.
@@ -3986,6 +4330,15 @@ class LlamaCppBackend:
                 "Install it with: pip install huggingface_hub"
             )
 
+        resolved_hf_repo = _resolve_repo_id_casing(hf_repo)
+        if resolved_hf_repo != hf_repo:
+            logger.info(
+                "Using cached repo_id casing '%s' for requested '%s'",
+                resolved_hf_repo,
+                hf_repo,
+            )
+            hf_repo = resolved_hf_repo
+
         # Resolve the filename from the variant
         gguf_filename = None
         gguf_extra_shards: list[str] = []
@@ -4031,10 +4384,12 @@ class LlamaCppBackend:
 
         # Check disk space; fall back to a smaller variant if needed
         all_gguf_files = [gguf_filename] + gguf_extra_shards
+        expected_sizes: dict[str, int] = {}
         try:
             from huggingface_hub import get_paths_info, try_to_load_from_cache
 
             path_infos = list(get_paths_info(hf_repo, all_gguf_files, token = hf_token))
+            expected_sizes = {p.path: p.size for p in path_infos if p.size}
             total_bytes = sum((p.size or 0) for p in path_infos)
 
             # Subtract bytes already in the HF cache so we only preflight
@@ -4043,7 +4398,26 @@ class LlamaCppBackend:
             # cold whenever free disk is below the full weight footprint,
             # even though nothing needs downloading.
             already_cached_bytes = 0
-            if not force:
+            # Cross-snapshot / case-variant cache reuse is offline-only (see the download
+            # path below); online, hf_hub_download fetches the current revision and
+            # resumes partials, so an old snapshot must not be counted as cached here or
+            # the preflight would under-count the download and skip the disk fallback.
+            offline = _hf_env_offline()
+            # A split GGUF whose shards are not co-located in a single snapshot is
+            # refetched as a whole set later, so it must not be counted as cached here.
+            split_needs_refetch = False
+            if offline and not force and gguf_extra_shards:
+                # Scan all snapshots for one that holds the whole set co-located, so a
+                # newer snapshot with only the first shard does not mask an older
+                # complete one and needlessly trip the disk fallback.
+                if (
+                    _cached_colocated_split_main(
+                        hf_repo, gguf_filename, gguf_extra_shards, expected_sizes
+                    )
+                    is None
+                ):
+                    split_needs_refetch = True
+            if not force and not split_needs_refetch:
                 for p in path_infos:
                     if not p.size:
                         continue
@@ -4051,6 +4425,15 @@ class LlamaCppBackend:
                         cached_path = try_to_load_from_cache(hf_repo, p.path)
                     except Exception:
                         cached_path = None
+                    if (
+                        not (isinstance(cached_path, str) and os.path.exists(cached_path))
+                        and offline
+                    ):
+                        cached_path = _cached_hf_snapshot_file(
+                            hf_repo,
+                            p.path,
+                            expected_size = p.size,
+                        )
                     if isinstance(cached_path, str) and os.path.exists(cached_path):
                         try:
                             on_disk = os.path.getsize(cached_path)
@@ -4113,6 +4496,13 @@ class LlamaCppBackend:
                             )
                         else:
                             gguf_extra_shards = []
+                        # Record the fallback's size so the later cache-reuse probe can
+                        # size-verify it; only for a single-file fallback, since
+                        # _find_smallest_fitting_variant returns the whole-variant size
+                        # and using that as the first shard's expected size would reject
+                        # a valid cached first shard of a split fallback.
+                        if not gguf_extra_shards:
+                            expected_sizes[fallback_file] = fallback_size
                     else:
                         raise RuntimeError(
                             f"Not enough disk space to download any variant. "
@@ -4132,25 +4522,45 @@ class LlamaCppBackend:
                 raise RuntimeError("Cancelled")
             dl_start = time.monotonic()
             # Xet primary, HTTP fallback on stall; per-file so finished shards stay cached.
-            local_path = hf_hub_download_with_xet_fallback(
-                hf_repo,
-                gguf_filename,
-                hf_token,
-                cancel_event = cancel_event,
-                on_status = lambda m: logger.info(m),
-                force_download = force,
-            )
-            for shard in gguf_extra_shards:
-                if cancel_event.is_set():
-                    raise RuntimeError("Cancelled")
-                logger.info(f"Resolving GGUF shard: {shard}")
-                hf_hub_download_with_xet_fallback(
+            local_path = None
+            # Reuse a cached copy from another snapshot / case-variant repo dir only when
+            # offline. Online, fall through to hf_hub_download so its revision/etag check
+            # fetches the current file (and resumes a partial) instead of serving a stale
+            # same-name blob from an older revision.
+            if not force and _hf_env_offline():
+                if gguf_extra_shards:
+                    # A split GGUF must load every shard from one snapshot; reuse only a
+                    # snapshot that holds the whole set co-located, scanning past a newer
+                    # snapshot that has just the first shard while an older one is complete.
+                    local_path = _cached_colocated_split_main(
+                        hf_repo, gguf_filename, gguf_extra_shards, expected_sizes
+                    )
+                else:
+                    local_path = _cached_hf_snapshot_file(
+                        hf_repo,
+                        gguf_filename,
+                        expected_size = expected_sizes.get(gguf_filename),
+                    )
+            if local_path is None:
+                local_path = hf_hub_download_with_xet_fallback(
                     hf_repo,
-                    shard,
+                    gguf_filename,
                     hf_token,
                     cancel_event = cancel_event,
+                    on_status = lambda m: logger.info(m),
                     force_download = force,
                 )
+                for shard in gguf_extra_shards:
+                    if cancel_event.is_set():
+                        raise RuntimeError("Cancelled")
+                    logger.info(f"Resolving GGUF shard: {shard}")
+                    hf_hub_download_with_xet_fallback(
+                        hf_repo,
+                        shard,
+                        hf_token,
+                        cancel_event = cancel_event,
+                        force_download = force,
+                    )
         except Exception as e:
             if isinstance(e, RuntimeError) and "Cancelled" in str(e):
                 raise
@@ -4228,6 +4638,17 @@ class LlamaCppBackend:
         if target is None or cancel_event.is_set():
             return None
 
+        # Offline, resolve the companion straight from the cache snapshot that
+        # holds it. resolve_cached_repo_id_case can return a partial lower-case
+        # spelling when any dir exists under the requested casing, so calling
+        # hf_hub_download with hf_repo would miss the canonical file and silently
+        # drop the companion. _cached_hf_snapshot_file scans every case variant.
+        if _hf_env_offline():
+            cached = _cached_hf_snapshot_file(hf_repo, target)
+            if cached:
+                logger.info("Resolved %s from local HF cache: %s", label, cached)
+                return cached
+
         try:
             logger.info(f"Downloading {label}: {hf_repo}/{target}")
             # Same policy; companions are best-effort (caller below swallows failures to None).
@@ -4276,6 +4697,29 @@ class LlamaCppBackend:
             cancel_event = cancel_event,
         )
 
+    def _cached_repo_mtp_drafter(self, hf_repo: str) -> Optional[str]:
+        """A drafter already in this repo's local HF cache, reused offline when a
+        fresh copy can't be fetched. Prefers a repo-root ``mtp-*.gguf`` across all
+        cached snapshots; else an existing ``MTP/`` copy (any precision -- the
+        target verifies every drafted token). None if none is cached."""
+        try:
+            from utils.models.model_config import _iter_hf_cache_snapshots
+
+            roots: list[Path] = []
+            subdirs: list[Path] = []
+            for snap in _iter_hf_cache_snapshots(hf_repo):  # newest first
+                for f in sorted(_gguf_snapshot_files(snap)):
+                    if _is_companion_gguf_path(f) and "mmproj" not in f.lower():
+                        (roots if "/" not in f else subdirs).append(snap / f)
+            # Keep snapshot order (newest first), root before any MTP/ copy, so a
+            # newer main GGUF pairs with the newest cached drafter, not a stale one.
+            for cand in roots + subdirs:
+                if cand.is_file():
+                    return str(cand)
+        except Exception as e:
+            logger.debug("Cached MTP drafter lookup failed for %s: %s", hf_repo, e)
+        return None
+
     def _download_mtp(
         self,
         *,
@@ -4292,11 +4736,25 @@ class LlamaCppBackend:
         are intentionally skipped. Returns the local path, or None.
         """
 
+        # Offline, reuse any drafter already on disk (a fresh copy can't be
+        # fetched). Online, _download_companion_gguf/hf_hub_download reuse the
+        # current cached file and refetch a changed one, so skip the probe here
+        # rather than pair new weights with a stale draft.
+        if _hf_env_offline():
+            cached = self._cached_repo_mtp_drafter(hf_repo)
+            if cached:
+                logger.info(f"Reusing cached MTP drafter (offline): {cached}")
+                return cached
+
         def _pick_mtp(candidates: list[str]) -> Optional[str]:
+            # Root-level only: MTP/ subdir copies now share the mtp- prefix but
+            # are explicit-selection, not auto-fetch (they'd sort ahead of root).
             mtp_files = sorted(
                 f
                 for f in candidates
-                if f.lower().endswith(".gguf") and Path(f).name.lower().startswith("mtp-")
+                if f.lower().endswith(".gguf")
+                and "/" not in f
+                and Path(f).name.lower().startswith("mtp-")
             )
             return mtp_files[0] if mtp_files else None
 
@@ -4998,6 +5456,7 @@ class LlamaCppBackend:
             # Resolve llama-server now but defer a not-found error: a block-diffusion
             # GGUF uses the diffusion runner, and its arch is only known after the header.
             binary = self._find_llama_server_binary()
+            is_vulkan_backend = self._is_vulkan_backend(binary)
 
             # ── Phase 2: download (NO lock held, so cancel can proceed) ──
             # mtp_draft_path arrives set for local Gemma loads (detected
@@ -5006,6 +5465,19 @@ class LlamaCppBackend:
             # dead; cleanup runs even on exception so a transient hiccup
             # can't quarantine future loads.
             if hf_repo:
+                # Resolve the requested repo id to its cached canonical casing once,
+                # up front, so the main GGUF and its companions (mmproj / MTP drafter)
+                # all resolve from the same cache entry. Otherwise a case-variant
+                # request resolves the main file from the canonical cache dir while the
+                # companions keep the requested casing and miss the cached files.
+                _resolved_repo = _resolve_repo_id_casing(hf_repo)
+                if _resolved_repo != hf_repo:
+                    logger.info(
+                        "Using cached repo_id casing '%s' for requested '%s'",
+                        _resolved_repo,
+                        hf_repo,
+                    )
+                    hf_repo = _resolved_repo
                 with _hf_offline_if_dns_dead():
                     model_path = self._download_gguf(
                         hf_repo = hf_repo,
@@ -5224,7 +5696,8 @@ class LlamaCppBackend:
                     model_size = gguf_size + mmproj_size
                     # 2-tuple gpus for existing logic + a total map for the absolute
                     # per-GPU headroom (correct when the GPU is already partly used).
-                    _gpu_mem = self._get_gpu_memory()
+                    # Pass binary so a Vulkan build probes ggml's Vulkan ordinals.
+                    _gpu_mem = self._get_gpu_memory(binary)
                     gpus = [(idx, free) for idx, free, _t in _gpu_mem]
                     total_by_idx = {idx: total for idx, _f, total in _gpu_mem}
 
@@ -5997,7 +6470,12 @@ class LlamaCppBackend:
                 # cap, not the ROCm-reported VRAM, is the real ceiling); refuse an
                 # oversize load the OS would otherwise kill mid-flight. Base model
                 # only: an optional MTP drafter is dropped by the MTP-drop fallback.
-                if model_size is not None and self._amd_apu_wants_unified_memory(gpu_indices):
+                # CUDA/ROCm ids only; a Vulkan build's gpu_indices are ggml ordinals.
+                if (
+                    model_size is not None
+                    and not is_vulkan_backend
+                    and self._amd_apu_wants_unified_memory(gpu_indices)
+                ):
                     _ram_msg = self._apu_ram_shortfall_message(
                         model_size, self._available_system_memory_mib()
                     )
@@ -6260,6 +6738,12 @@ class LlamaCppBackend:
                             ", ".join(unsupported_cache_flags),
                         )
 
+                # Vulkan pins via --device (a cmd arg, unlike the env-based
+                # CUDA/ROCm pin below), emitted BEFORE user extras so llama.cpp's
+                # last-wins parsing lets a user --device override Studio's pick.
+                if is_vulkan_backend and gpu_indices is not None:
+                    cmd += LlamaCppBackend._vulkan_pin_args(gpu_indices)
+
                 # User pass-through args go last so llama.cpp's last-wins parsing
                 # lets the user override Studio's auto-set flags. Already
                 # validated by the route via validate_extra_args().
@@ -6311,23 +6795,25 @@ class LlamaCppBackend:
                         env.setdefault("OMP_NUM_THREADS", "2")
 
                 # AMD unified-memory APUs (gfx1150/gfx1151): let llama.cpp use
-                # shared system RAM. setdefault so a user value wins.
-                if self._amd_apu_wants_unified_memory(gpu_indices):
+                # shared system RAM. setdefault so a user value wins. Not on Vulkan
+                # (nor DC below): gpu_indices are ggml ordinals, not CUDA/ROCm ids.
+                if not is_vulkan_backend and self._amd_apu_wants_unified_memory(gpu_indices):
                     env.setdefault("GGML_CUDA_ENABLE_UNIFIED_MEMORY", "1")
                     logger.info("AMD unified-memory APU: set GGML_CUDA_ENABLE_UNIFIED_MEMORY=1")
 
                 # DC NVIDIA GPUs: FP32 accum (+ P2P / launch queues for multi-GPU).
                 # See _apply_datacenter_env; opt out with UNSLOTH_DISABLE_DC_TUNING=1.
-                if self._apply_datacenter_env(env, gpu_indices):
+                if not is_vulkan_backend and self._apply_datacenter_env(env, gpu_indices):
                     multi_gpu = self._effective_gpu_count(gpu_indices) > 1
                     logger.info(
                         f"Data-center GPU detected: applied DC llama.cpp env tuning (multi_gpu={multi_gpu})"
                     )
 
                 # Pin to selected GPU(s). On ROCm, narrowing only
-                # CUDA_VISIBLE_DEVICES leaves an AMD child seeing the full
-                # set, so set HIP_VISIBLE_DEVICES too.
-                if gpu_indices is not None:
+                # CUDA_VISIBLE_DEVICES leaves an AMD child seeing the full set, so
+                # set HIP_VISIBLE_DEVICES too. Vulkan is pinned via --device
+                # (above), not here.
+                if gpu_indices is not None and not is_vulkan_backend:
                     pinned = ",".join(str(i) for i in gpu_indices)
                     env["CUDA_VISIBLE_DEVICES"] = pinned
                     try:
@@ -6721,6 +7207,7 @@ class LlamaCppBackend:
                         )
 
                 self._healthy = True
+                self._commit_effective_parallel_slots(n_parallel)
 
                 # Commit caller intent only after _healthy=True so a failed start
                 # can't poison the next inheritance check. None keeps prior, []
@@ -7258,6 +7745,7 @@ class LlamaCppBackend:
             self._context_length = None
             self._effective_context_length = None
             self._max_context_length = None
+            self._reset_effective_parallel_slots()
             self._chat_template = None
             self._chat_template_override = None
             self._supports_reasoning = False
@@ -7313,6 +7801,7 @@ class LlamaCppBackend:
         # Stop the watchdog before a deliberate kill so a planned reload/unload
         # isn't seen as a crash; a real crash never routes through here.
         self._stop_mtp_crash_watchdog()
+        self._reset_effective_parallel_slots()
         if self._process is None:
             return
         try:
@@ -8154,7 +8643,7 @@ class LlamaCppBackend:
     ):
         """Open one streaming POST and let cancel interrupt prefill or reads."""
         if cancel_event is not None and cancel_event.is_set():
-            raise GeneratorExit
+            raise _LlamaStreamCancelled
 
         _cancel_closed = threading.Event()
         _response_ref: list = [None]
@@ -8199,13 +8688,13 @@ class LlamaCppBackend:
             ) as response:
                 _response_ref[0] = response
                 if cancel_event is not None and cancel_event.is_set():
-                    raise GeneratorExit
+                    raise _LlamaStreamCancelled
                 yield response
                 return
         except (httpx.RequestError, RuntimeError):
             # Response was closed by the cancel watcher
             if cancel_event is not None and cancel_event.is_set():
-                raise GeneratorExit
+                raise _LlamaStreamCancelled
             raise
         finally:
             _cancel_closed.set()
@@ -8408,6 +8897,8 @@ class LlamaCppBackend:
                         "finish_reason": _metadata_finish_reason,
                     }
 
+        except _LlamaStreamCancelled:
+            return
         except httpx.ConnectError as e:
             # Server already down. If this was an MTP+tensor crash, recover by
             # reloading without MTP (scheduled in the background) and fail this
@@ -9532,6 +10023,8 @@ class LlamaCppBackend:
                         break
                 continue
 
+            except _LlamaStreamCancelled:
+                return
             except httpx.ConnectError:
                 # Mark unresolved provisional cards as failed before raising.
                 for _pid, _pname in provisional_started_tool_calls.items():
@@ -9714,6 +10207,8 @@ class LlamaCppBackend:
                 if _meta is not None:
                     yield _meta
 
+        except _LlamaStreamCancelled:
+            return
         except httpx.ConnectError:
             raise RuntimeError("Lost connection to llama-server")
         except Exception as e:
diff --git a/studio/backend/core/inference/mlx_inference.py b/studio/backend/core/inference/mlx_inference.py
index 62d268e15f..6287b184a6 100644
--- a/studio/backend/core/inference/mlx_inference.py
+++ b/studio/backend/core/inference/mlx_inference.py
@@ -5,6 +5,7 @@ Drop-in replacement for InferenceBackend — same interface, uses mlx-lm/mlx-vlm
 instead of torch/transformers for model loading and generation.
 """
 
+import os
 import threading
 from typing import Optional, Generator
 from core.inference.runtime_context import runtime_context_length
@@ -41,6 +42,48 @@ def _build_generation_stats(prompt_n, prompt_tps, gen_n, gen_tps):
     }
 
 
+def _mlx_distributed_rank_size(group = None):
+    """Return ``(rank, world_size)`` for an optional MLX distributed group."""
+    if group is None:
+        return 0, 1
+    rank = int(group.rank())
+    world_size = int(group.size())
+    if world_size < 1:
+        raise ValueError(f"Invalid MLX distributed world_size={world_size}.")
+    if rank < 0 or rank >= world_size:
+        raise ValueError(f"Invalid MLX distributed rank={rank} for world_size={world_size}.")
+    return rank, world_size
+
+
+def _mlx_distributed_backend_from_env():
+    if os.environ.get("MLX_JACCL_COORDINATOR") and os.environ.get("MLX_IBV_DEVICES"):
+        return "jaccl"
+    return None
+
+
+def _init_mlx_distributed():
+    """Initialize MLX distributed state, falling back to singleton metadata."""
+    import mlx.core as mx
+
+    group = None
+    rank = 0
+    world_size = 1
+    distributed = getattr(mx, "distributed", None)
+    init = getattr(distributed, "init", None) if distributed is not None else None
+    if callable(init):
+        backend = _mlx_distributed_backend_from_env()
+        if backend is None:
+            group = init()
+        else:
+            try:
+                group = init(backend = backend)
+            except TypeError:
+                group = init()
+        if group is not None:
+            rank, world_size = _mlx_distributed_rank_size(group)
+    return group, rank, world_size
+
+
 def _make_mlx_presence_penalty_processor(penalty: float):
     """Presence penalty as an mlx_lm/mlx_vlm logits processor, matching the safetensors path.
 
@@ -52,7 +95,7 @@ def _make_mlx_presence_penalty_processor(penalty: float):
 
     def _processor(tokens, logits):
         if state["prompt_len"] is None:
-            # First call = prompt only; latch its length.
+            # First call is prompt-only; latch its length.
             state["prompt_len"] = int(tokens.shape[0])
             return logits
         generated = tokens[state["prompt_len"] :]
@@ -61,22 +104,17 @@ def _make_mlx_presence_penalty_processor(penalty: float):
         import mlx.core as mx
 
         vocab = logits.shape[-1]
-        # Bound generated ids to the valid range [0, vocab) before they index
-        # logits. MLX does no bounds checking and out-of-bounds indexing is
-        # documented undefined behavior (crash / memory corruption), unlike the
-        # torch path's harmless negative wrap -- so this bound is load-bearing
-        # here and matches the torch filter seen[(seen >= 0) & (seen < vocab)].
-        # MLX has no boolean-mask filtering (data-dependent output shape is
-        # unsupported), so instead of compacting the id list we route every
-        # out-of-range or negative id to a scratch slot at index ``vocab`` that
-        # is dropped before the subtract. That scratch slot can never collide
-        # with a real token, so real ids (including id 0) are penalized exactly
-        # once and stray ids are ignored.
+        # Bound ids to [0, vocab) before indexing logits: MLX does no bounds
+        # checking and out-of-bounds indexing is undefined behavior (crash /
+        # corruption), unlike torch's harmless negative wrap. MLX also lacks
+        # boolean-mask filtering, so out-of-range/negative ids route to a
+        # scratch slot at index vocab (dropped before the subtract) that never
+        # collides with a real token: real ids (including 0) are penalized
+        # once, strays ignored.
         valid = (generated >= 0) & (generated < vocab)
         safe = mx.where(valid, generated, vocab).astype(mx.int32)
-        # Scatter-assign a scalar penalty into a (vocab + 1)-wide mask: duplicate
-        # ids are idempotent, so presence applies once per distinct token; the
-        # scratch column is discarded and the full-width subtract stays on-device.
+        # Scatter penalty into a (vocab + 1)-wide mask: duplicate ids are
+        # idempotent (presence applies once per token); scratch column dropped.
         mask = mx.zeros((vocab + 1,), dtype = logits.dtype)
         mask[safe] = penalty
         logits = logits - mask[:vocab]
@@ -93,7 +131,7 @@ class MLXInferenceBackend:
         self.loaded_local_models = []
         self.device = "mlx"
         self._generation_lock = threading.Lock()
-        # usage/timings of the latest generation; shipped on gen_done.
+        # usage/timings of the latest generation, shipped on gen_done.
         self.last_generation_stats = None
 
         self._model = None
@@ -101,6 +139,9 @@ class MLXInferenceBackend:
         self._processor = None
         self._is_vlm = False
         self._config = {}
+        self._distributed_group = None
+        self._distributed_rank = 0
+        self._distributed_world_size = 1
 
         # Recorded for unload to release pinned memory back to the OS.
         self._memory_limits_applied = {}
@@ -145,19 +186,26 @@ class MLXInferenceBackend:
         trust_remote_code = False,
         gpu_ids = None,
         dtype = None,
+        parallel_mode = None,
+        distributed_group = None,
     ) -> bool:
         import mlx.core as mx
 
-        # Keep the token so the native-template fallback can fetch a
-        # gated model's repo template later during generation.
+        # Keep the token so the native-template fallback can fetch a gated
+        # model's repo template during generation.
         self._hf_token = hf_token
         model_name = config.identifier if hasattr(config, "identifier") else str(config)
         is_vision = getattr(config, "is_vision", False)
+        distributed_rank, distributed_size = _mlx_distributed_rank_size(distributed_group)
+        is_distributed = distributed_group is not None and distributed_size > 1
+        self._distributed_group = distributed_group
+        self._distributed_rank = distributed_rank
+        self._distributed_world_size = distributed_size
 
-        # GGUF guard. GGUF models are served by llama-server in the parent
-        # process, not mlx-lm here. Reaching this with is_gguf=True means the
-        # route's first detection flaked (transient HF Hub) but the subprocess
-        # re-detected GGUF; raise loudly instead of a cryptic mlx_lm error.
+        # GGUF guard: GGUF is served by llama-server in the parent process,
+        # not mlx-lm. Reaching here with is_gguf=True means the route's
+        # detection flaked but the subprocess re-detected GGUF; raise loudly
+        # instead of a cryptic mlx_lm error.
         if getattr(config, "is_gguf", False):
             raise RuntimeError(
                 f"MLXInferenceBackend cannot load GGUF model '{model_name}': "
@@ -176,11 +224,26 @@ class MLXInferenceBackend:
         is_lora = getattr(config, "is_lora", False)
 
         logger.info(
-            "Loading %s via %s (is_lora=%s)",
+            "Loading %s via %s (is_lora=%s, distributed=%s, rank=%s/%s, mode=%s)",
             model_name,
             "mlx-vlm" if is_vision else "mlx-lm",
             is_lora,
+            is_distributed,
+            distributed_rank,
+            distributed_size,
+            parallel_mode,
         )
+        if is_distributed and parallel_mode not in ("pipeline", "tensor"):
+            raise ValueError(
+                "Unsloth: distributed MLX inference requires parallel_mode='pipeline' "
+                "or parallel_mode='tensor'."
+            )
+        if is_distributed and is_lora:
+            raise ValueError(
+                "Unsloth: distributed MLX inference for LoRA adapter repos "
+                "is not supported yet. Merge/export the adapter into an MLX model "
+                "before distributed inference."
+            )
 
         try:
             from unsloth_zoo.mlx.loader import FastMLXModel
@@ -190,14 +253,23 @@ class MLXInferenceBackend:
                 "(unsloth_zoo.mlx.loader). Reinstall via install.sh on Apple Silicon."
             ) from e
 
+        load_kwargs = {
+            "max_seq_length": max_seq_length,
+            "dtype": dtype,
+            "load_in_4bit": load_in_4bit,
+            "token": hf_token,
+            "trust_remote_code": trust_remote_code,
+            "text_only": False if is_vision else True,
+        }
+        if is_distributed:
+            if parallel_mode == "pipeline":
+                load_kwargs["pipeline_group"] = distributed_group
+            else:
+                load_kwargs["tensor_group"] = distributed_group
+
         model, tokenizer_or_processor = FastMLXModel.from_pretrained(
             model_name,
-            max_seq_length = max_seq_length,
-            dtype = dtype,
-            load_in_4bit = load_in_4bit,
-            token = hf_token,
-            trust_remote_code = trust_remote_code,
-            text_only = False if is_vision else True,
+            **load_kwargs,
         )
 
         if is_vision:
@@ -217,8 +289,7 @@ class MLXInferenceBackend:
         self.models[model_name] = {
             # Per-model token for the native-template fallback (matches transformers).
             "hf_token": hf_token,
-            # Per-model consent for the native-template reload: re-use the exact
-            # trust_remote_code this model was loaded with (matches transformers).
+            # Per-model trust_remote_code reused by the native-template reload (matches transformers).
             "trust_remote_code": trust_remote_code,
             "model": self._model,
             "tokenizer": self._tokenizer,
@@ -234,8 +305,7 @@ class MLXInferenceBackend:
             "has_audio_input": False,
             "context_length": runtime_context_length(self._model, max_seq_length),
         }
-        # Capture chat_template_info so the worker IPC reply ships it back and
-        # the route layer classifies capabilities like the other paths.
+        # Capture chat_template_info for the worker IPC reply and route capability classification.
         self._populate_chat_template_info(model_name)
 
         logger.info("Model %s loaded successfully", model_name)
@@ -293,6 +363,9 @@ class MLXInferenceBackend:
         self._model = None
         self._tokenizer = None
         self._processor = None
+        self._distributed_group = None
+        self._distributed_rank = 0
+        self._distributed_world_size = 1
         if self.active_model_name == model_name:
             self.active_model_name = None
         gc.collect()
@@ -320,8 +393,7 @@ class MLXInferenceBackend:
         max_new_tokens = 256,
         repetition_penalty = 1.0,
         cancel_event = None,
-        # Reasoning / tool kwargs forwarded by the route + worker; rendered via
-        # apply_chat_template_for_generation like the transformers path.
+        # Reasoning / tool kwargs, rendered via apply_chat_template_for_generation (transformers parity).
         tools = None,
         enable_thinking = None,
         reasoning_effort = None,
@@ -334,7 +406,6 @@ class MLXInferenceBackend:
         # Reset so a failed run cannot surface stale stats.
         self.last_generation_stats = None
 
-        # Build messages with system prompt
         full_messages = []
         if system_prompt:
             full_messages.append({"role": "system", "content": system_prompt})
@@ -351,7 +422,6 @@ class MLXInferenceBackend:
                             {"type": "text", "text": content},
                         ]
                     elif isinstance(content, list):
-                        # Prepend image if not already present
                         has_image = any(
                             p.get("type") == "image" for p in content if isinstance(p, dict)
                         )
@@ -415,6 +485,7 @@ class MLXInferenceBackend:
 
         from core.inference.chat_template_helpers import (
             apply_chat_template_for_generation,
+            detect_think_prefill,
             render_with_native_template_fallback,
         )
 
@@ -429,11 +500,11 @@ class MLXInferenceBackend:
         if prompt is None:
             raise RuntimeError("apply_chat_template returned None — tokenizer may be incompatible")
 
-        # Same parity fix as the transformers backend: if the template dropped the
-        # requested tools, fall back to the native template so MLX text models keep
-        # advertising them. ``self._tokenizer`` is this entry's model_info tokenizer,
-        # so probe and native render share a renderer. (The VLM path renders via the
-        # processor for image tokens and is intentionally not wired here.)
+        # Parity with the transformers backend: if the template dropped the
+        # requested tools, fall back to the native template so MLX text models
+        # keep advertising them. self._tokenizer is this entry's tokenizer, so
+        # probe and native render share a renderer. (VLM renders via the
+        # processor for image tokens and is not wired here.)
         model_info = self.models.get(self.active_model_name, {})
         prompt = render_with_native_template_fallback(
             formatted_prompt = prompt,
@@ -448,6 +519,15 @@ class MLXInferenceBackend:
             hf_token = model_info.get("hf_token"),
         )
 
+        # An open <think> prefilled by the template lives in the prompt, not
+        # the generated tokens; re-emit it so the frontend renders the block.
+        think_prefix = detect_think_prefill(
+            prompt, getattr(self._tokenizer, "all_special_tokens", None)
+        )
+        # Emit it before the first token so the block renders during prefill.
+        if think_prefix:
+            yield think_prefix
+
         sampler = make_sampler(
             temp = temperature,
             top_p = top_p,
@@ -455,7 +535,7 @@ class MLXInferenceBackend:
             min_p = float(min_p or 0.0),
             min_tokens_to_keep = 1,
         )
-        # Repetition and/or presence penalty processors (parity with the GGUF/safetensors paths).
+        # Repetition and/or presence penalty processors (GGUF/safetensors parity).
         logits_processors = []
         if repetition_penalty is not None and float(repetition_penalty) not in (
             0.0,
@@ -496,12 +576,11 @@ class MLXInferenceBackend:
                 ):
                     final_response = response
                     token_ids.append(response.token)
-                    # Decode full sequence with skip_special_tokens
                     cumulative = self._tokenizer.decode(
                         token_ids,
                         skip_special_tokens = True,
                     )
-                    yield cumulative
+                    yield think_prefix + cumulative
 
                     if cancel_event and cancel_event.is_set():
                         break
@@ -544,8 +623,7 @@ class MLXInferenceBackend:
         )
 
         # Pick the chat-template-aware caller: processors with their own
-        # apply_chat_template + chat_template (e.g. Qwen2.5-VL) use it
-        # directly; else fall back to the nested tokenizer.
+        # apply_chat_template + chat_template (e.g. Qwen2.5-VL), else the nested tokenizer.
         chat_target = self._processor
         if (
             getattr(self._processor, "apply_chat_template", None) is None
@@ -566,16 +644,21 @@ class MLXInferenceBackend:
         # mlx_vlm's stream_generate handles pixel_values (None for text-only)
         images = [image] if image is not None else None
 
-        cumulative = ""
+        from core.inference.chat_template_helpers import detect_think_prefill
+
+        # Re-emit an open <think> prefill from the prompt (see _generate_text).
+        cumulative = detect_think_prefill(prompt, getattr(chat_target, "all_special_tokens", None))
+        # Emit it before the first token so the block renders during prefill.
+        if cumulative:
+            yield cumulative
         logger.info(
             "VLM generating: prompt_len=%d, has_image=%s",
             len(prompt),
             image is not None,
         )
-        # mlx_vlm.stream_generate forwards **kwargs into generate_step, which
-        # builds the sampler + logits_processors internally.
-        # GOTCHA: generate_step expects ``temperature=`` (long form); ``temp=``
-        # silently falls into **kwargs and is ignored, stuck at greedy 0.0.
+        # stream_generate forwards **kwargs into generate_step (builds the
+        # sampler + logits_processors internally). GOTCHA: generate_step expects
+        # temperature= (long form); temp= is silently ignored, stuck at greedy 0.0.
         vlm_kwargs = dict(
             max_tokens = max_new_tokens,
             temperature = temperature,
@@ -589,7 +672,7 @@ class MLXInferenceBackend:
         )
         if presence_penalty:
             # Presence needs a custom processor: pass the full list (repetition +
-            # presence) instead of the repetition_penalty shortcut so both apply once.
+            # presence) instead of the repetition_penalty shortcut so both apply.
             from mlx_lm.sample_utils import make_logits_processors
 
             _vlm_processors = []
@@ -634,7 +717,7 @@ class MLXInferenceBackend:
         cancel_event = None,
         **gen_kwargs,
     ) -> Generator[str, None, None]:
-        # MLX LoRA adapter toggling not yet supported — generate normally
+        # MLX LoRA adapter toggling not yet supported; generate normally
         yield from self.generate_chat_response(cancel_event = cancel_event, **gen_kwargs)
 
     def reset_generation_state(self):
diff --git a/studio/backend/core/inference/orchestrator.py b/studio/backend/core/inference/orchestrator.py
index cf5d24c367..4fa0d3ed26 100644
--- a/studio/backend/core/inference/orchestrator.py
+++ b/studio/backend/core/inference/orchestrator.py
@@ -50,6 +50,18 @@ _DISPATCH_DRAIN_TIMEOUT = 5.0
 _UNLOAD_GEN_LOCK_TIMEOUT = 15.0
 
 
+class GenStreamError(str):
+    """A stream chunk carrying a real backend/generation error, not model text.
+
+    Subclasses str so existing display/logging consumers are unaffected, while
+    callers that must abort a distributed run on error (raise_on_streamed_error)
+    can distinguish a real error from model output whose visible text starts with
+    "Error:" by checking isinstance(chunk, GenStreamError).
+    """
+
+    __slots__ = ()
+
+
 class InferenceOrchestrator:
     """
     Inference backend orchestrator — subprocess-based.
@@ -482,13 +494,13 @@ class InferenceOrchestrator:
         initial_resp_queue = self._resp_queue
         while True:
             if self._proc is not initial_proc or self._resp_queue is not initial_resp_queue:
-                yield f"Error: {self._subprocess_crash_message(crash_context)}"
+                yield GenStreamError(f"Error: {self._subprocess_crash_message(crash_context)}")
                 return
             resp = read_one(read_timeout)
             if resp is None:
                 # Check subprocess health
                 if not self._ensure_subprocess_alive():
-                    yield f"Error: {self._subprocess_crash_message(crash_context)}"
+                    yield GenStreamError(f"Error: {self._subprocess_crash_message(crash_context)}")
                     return
                 continue
 
@@ -498,7 +510,7 @@ class InferenceOrchestrator:
             # Subprocess-level error (no request_id); request-scoped failures
             # arrive as gen_error below.
             if rtype == "error" and not resp.get("request_id"):
-                yield f"Error: {resp.get('error', 'Unknown error')}"
+                yield GenStreamError(f"Error: {resp.get('error', 'Unknown error')}")
                 return
 
             if rtype == "token":
@@ -513,7 +525,7 @@ class InferenceOrchestrator:
                     stats_holder["stats"] = resp.get("stats")
                 return
             elif rtype == "gen_error":
-                yield f"Error: {resp.get('error', 'Unknown error')}"
+                yield GenStreamError(f"Error: {resp.get('error', 'Unknown error')}")
                 return
 
     # ------------------------------------------------------------------
@@ -640,11 +652,11 @@ class InferenceOrchestrator:
         GPU work stays serialized; this only avoids orchestrator lock contention.
         """
         if not self._ensure_subprocess_alive():
-            yield "Error: Inference subprocess is not running"
+            yield GenStreamError("Error: Inference subprocess is not running")
             return
 
         if not self.active_model_name:
-            yield "Error: No active model"
+            yield GenStreamError("Error: No active model")
             return
         # Latch the target model so the recheck below can detect a switch that completed
         # between _start_dispatcher and mailbox registration (mirrors the locked path's
@@ -655,7 +667,7 @@ class InferenceOrchestrator:
         # so without this early-out a compare request would enqueue a generate on the
         # outgoing model and delay the switch.
         if self._unload_pending:
-            yield "Error: model is being unloaded"
+            yield GenStreamError("Error: model is being unloaded")
             return
 
         # Ensure the dispatcher runs. _start_dispatcher serializes concurrent starters under
@@ -727,7 +739,7 @@ class InferenceOrchestrator:
             # _stop_dispatcher joins the dispatcher, which itself takes that lock.
             if orphaned_dispatcher:
                 self._stop_dispatcher()
-            yield "Error: model is being unloaded"
+            yield GenStreamError("Error: model is being unloaded")
             return
 
         try:
@@ -735,7 +747,7 @@ class InferenceOrchestrator:
         except RuntimeError as exc:
             with self._mailbox_lock:
                 self._mailboxes.pop(request_id, None)
-            yield f"Error: {exc}"
+            yield GenStreamError(f"Error: {exc}")
             return
 
         def read_mailbox(timeout):
@@ -813,6 +825,59 @@ class InferenceOrchestrator:
         self._stop_dispatcher()
         return True
 
+    def share_distributed_object(
+        self,
+        obj,
+        timeout: Optional[float] = 300.0,
+    ):
+        """Share a small object through the worker's MLX distributed group."""
+        if not self._ensure_subprocess_alive():
+            raise RuntimeError("Inference subprocess is not running")
+
+        self._wait_dispatcher_idle()
+        with self._mailbox_lock:
+            if self._mailboxes:
+                raise RuntimeError(
+                    "Cannot share distributed objects while compare requests are active"
+                )
+        request_id = str(uuid.uuid4())
+        cmd = {
+            "type": "share_object",
+            "request_id": request_id,
+            "object": obj,
+        }
+
+        with self._gen_lock:
+            self._send_cmd(cmd)
+            deadline = None if timeout is None else time.monotonic() + timeout
+            while deadline is None or time.monotonic() < deadline:
+                remaining = 1.0 if deadline is None else max(0.1, deadline - time.monotonic())
+                resp = self._read_resp(timeout = min(remaining, 1.0))
+                if resp is None:
+                    if not self._ensure_subprocess_alive():
+                        raise RuntimeError(self._subprocess_crash_message("sharing chat turn"))
+                    continue
+
+                rtype = resp.get("type", "")
+                rid = resp.get("request_id")
+                if rid and rid != request_id:
+                    logger.debug(
+                        "Skipping response for request_id=%s while sharing request_id=%s",
+                        rid,
+                        request_id,
+                    )
+                    continue
+                if rtype == "shared":
+                    return resp.get("object")
+                if rtype == "share_error":
+                    raise RuntimeError(resp.get("error", "Failed to share object"))
+                if rtype == "error":
+                    raise RuntimeError(resp.get("error", "Subprocess error"))
+                if rtype == "status":
+                    continue
+
+            raise RuntimeError("Timeout waiting for distributed object share")
+
     # ------------------------------------------------------------------
     # Public API — same interface as InferenceBackend
     # ------------------------------------------------------------------
@@ -828,6 +893,8 @@ class InferenceOrchestrator:
         approved_remote_code_fingerprint: Optional[str] = None,
         gpu_ids: Optional[list[int]] = None,
         subject: Optional[str] = None,
+        tensor_parallel: bool = False,
+        mlx_distributed: bool = False,
     ) -> bool:
         """Load a model for inference.
 
@@ -853,6 +920,11 @@ class InferenceOrchestrator:
                 "approved_remote_code_fingerprint": approved_remote_code_fingerprint,
                 "subject": subject,
                 "gpu_ids": gpu_ids,
+                "tensor_parallel": bool(tensor_parallel),
+                "mlx_distributed": bool(mlx_distributed),
+                "mlx_parallel_mode": ("tensor" if tensor_parallel else "pipeline")
+                if mlx_distributed
+                else None,
             }
             resolved_gpu_ids, gpu_selection = prepare_gpu_selection(
                 gpu_ids,
@@ -1338,11 +1410,11 @@ class InferenceOrchestrator:
         readers don't consume each other's tokens off the shared resp_queue.
         """
         if not self._ensure_subprocess_alive():
-            yield "Error: Inference subprocess is not running"
+            yield GenStreamError("Error: Inference subprocess is not running")
             return
 
         if not self.active_model_name:
-            yield "Error: No active model"
+            yield GenStreamError("Error: No active model")
             return
         expected_model = self.active_model_name
 
@@ -1359,7 +1431,7 @@ class InferenceOrchestrator:
             # so we never generate on the wrong one.
             if self._unload_pending or self.active_model_name != expected_model:
                 # Won the lock handoff during a switch; don't start on the outgoing model.
-                yield "Error: model is being unloaded"
+                yield GenStreamError("Error: model is being unloaded")
                 return
             request_id = str(uuid.uuid4())
             image_b64 = self._pil_to_base64(image) if image is not None else None
@@ -1385,7 +1457,7 @@ class InferenceOrchestrator:
             try:
                 self._send_cmd(cmd)
             except RuntimeError as exc:
-                yield f"Error: {exc}"
+                yield GenStreamError(f"Error: {exc}")
                 return
 
             yield from self._consume_token_stream(
@@ -1544,10 +1616,10 @@ class InferenceOrchestrator:
     ) -> Generator[str, None, None]:
         """Shared inner logic for audio input generation (Whisper + ASR)."""
         if not self._ensure_subprocess_alive():
-            yield "Error: Inference subprocess is not running"
+            yield GenStreamError("Error: Inference subprocess is not running")
             return
         if not self.active_model_name:
-            yield "Error: No active model"
+            yield GenStreamError("Error: No active model")
             return
         expected_model = self.active_model_name
 
@@ -1556,7 +1628,7 @@ class InferenceOrchestrator:
             # cleared or swapped the model while we waited.
             if self._unload_pending or self.active_model_name != expected_model:
                 # Won the lock handoff during a switch; don't start on the outgoing model.
-                yield "Error: model is being unloaded"
+                yield GenStreamError("Error: model is being unloaded")
                 return
             request_id = str(uuid.uuid4())
 
@@ -1583,7 +1655,7 @@ class InferenceOrchestrator:
             try:
                 self._send_cmd(cmd)
             except RuntimeError as exc:
-                yield f"Error: {exc}"
+                yield GenStreamError(f"Error: {exc}")
                 return
 
             yield from self._consume_token_stream(
diff --git a/studio/backend/core/inference/worker.py b/studio/backend/core/inference/worker.py
index d4b102e422..05dee39283 100644
--- a/studio/backend/core/inference/worker.py
+++ b/studio/backend/core/inference/worker.py
@@ -13,6 +13,7 @@ mp.Queue, and exits on shutdown or unload. Pattern follows core/training/worker.
 from __future__ import annotations
 
 import base64
+import json
 from loggers import get_logger
 import os
 import queue as _queue
@@ -26,6 +27,9 @@ from typing import Any
 logger = get_logger(__name__)
 from utils.hardware import apply_gpu_ids
 
+_SHARE_OBJECT_MAX_BYTES = 1 << 20
+_SHARE_OBJECT_ERROR_SIZE = -1
+
 # studio/backend root, prepended to sys.path so the spawned subprocess can
 # import the utils/core packages.
 _BACKEND_PATH = str(Path(__file__).resolve().parent.parent.parent)
@@ -75,6 +79,17 @@ def _send_response(resp_queue: Any, response: dict) -> None:
         logger.error("Failed to send response: %s", exc)
 
 
+def _encode_share_object(obj: Any) -> bytes:
+    data = json.dumps(obj, separators = (",", ":"), ensure_ascii = False).encode("utf-8")
+    if len(data) > _SHARE_OBJECT_MAX_BYTES:
+        raise ValueError("Distributed object share payload is too large")
+    return data
+
+
+def _decode_share_object(data: Any) -> Any:
+    return json.loads(bytes(data.tolist()).decode("utf-8"))
+
+
 def _clean_token(value: str | None) -> str | None:
     """Normalize an HF token: blank or whitespace-only becomes None."""
     return value if value and value.strip() else None
@@ -329,14 +344,18 @@ def _handle_load(backend, config: dict, resp_queue: Any) -> None:
             xet_disabled = os.environ.get("HF_HUB_DISABLE_XET") == "1",
         )
         try:
-            success = backend.load_model(
-                config = mc,
-                max_seq_length = config.get("max_seq_length", 2048),
-                load_in_4bit = load_in_4bit,
-                hf_token = hf_token,
-                trust_remote_code = trust_remote_code,
-                gpu_ids = config.get("resolved_gpu_ids"),
-            )
+            load_kwargs = {
+                "config": mc,
+                "max_seq_length": config.get("max_seq_length", 2048),
+                "load_in_4bit": load_in_4bit,
+                "hf_token": hf_token,
+                "trust_remote_code": trust_remote_code,
+                "gpu_ids": config.get("resolved_gpu_ids"),
+            }
+            if getattr(backend, "device", None) == "mlx":
+                load_kwargs["parallel_mode"] = config.get("mlx_parallel_mode")
+                load_kwargs["distributed_group"] = config.get("_mlx_distributed_group")
+            success = backend.load_model(**load_kwargs)
         finally:
             heartbeat_stop.set()
 
@@ -521,6 +540,67 @@ def _handle_generate(backend, cmd: dict, resp_queue: Any, cancel_event) -> None:
         )
 
 
+def _handle_share_object(backend, cmd: dict, resp_queue: Any) -> None:
+    """Share a small Python object across MLX distributed ranks."""
+    request_id = cmd.get("request_id", "")
+    group = getattr(backend, "_distributed_group", None)
+    rank = int(getattr(backend, "_distributed_rank", 0) or 0)
+    world_size = int(getattr(backend, "_distributed_world_size", 1) or 1)
+    obj = cmd.get("object")
+
+    try:
+        if group is None or world_size <= 1:
+            shared = obj
+        else:
+            import mlx.core as mx
+            if rank == 0:
+                if obj is None:
+                    mx.eval(mx.distributed.all_sum(mx.array(0), group = group))
+                    shared = None
+                else:
+                    try:
+                        data = mx.array(_encode_share_object(obj), dtype = mx.uint8)
+                    except Exception:
+                        mx.eval(
+                            mx.distributed.all_sum(
+                                mx.array(_SHARE_OBJECT_ERROR_SIZE),
+                                group = group,
+                            )
+                        )
+                        raise
+                    mx.eval(mx.distributed.all_sum(mx.array(data.size), group = group))
+                    mx.eval(mx.distributed.all_sum(data, group = group))
+                    shared = obj
+            else:
+                size = int(mx.distributed.all_sum(mx.array(0), group = group).item())
+                if size == _SHARE_OBJECT_ERROR_SIZE:
+                    raise RuntimeError("Failed to share distributed object")
+                if size == 0:
+                    shared = None
+                else:
+                    data = mx.zeros(size, dtype = mx.uint8)
+                    data = mx.distributed.all_sum(data, group = group)
+                    shared = _decode_share_object(data)
+        _send_response(
+            resp_queue,
+            {
+                "type": "shared",
+                "request_id": request_id,
+                "object": shared,
+            },
+        )
+    except Exception as exc:
+        _send_response(
+            resp_queue,
+            {
+                "type": "share_error",
+                "request_id": request_id,
+                "error": str(exc),
+                "stack": traceback.format_exc(limit = 20),
+            },
+        )
+
+
 def _handle_generate_audio(backend, cmd: dict, resp_queue: Any) -> None:
     """Handle TTS audio generation — returns WAV bytes + sample_rate."""
     request_id = cmd.get("request_id", "")
@@ -720,9 +800,29 @@ def run_inference_process(
                 exc,
             )
         try:
-            from core.inference.mlx_inference import MLXInferenceBackend
+            from core.inference.mlx_inference import MLXInferenceBackend, _init_mlx_distributed
 
             backend = MLXInferenceBackend()
+            if config.get("mlx_distributed"):
+                group, rank, size = _init_mlx_distributed()
+                config["_mlx_distributed_group"] = group
+                if size <= 1:
+                    # A singleton group (MLX built without distributed support,
+                    # or an invalid launch env/hostfile) would leave nonzero ranks
+                    # looping forever on share_distributed_object. Fail the load
+                    # instead of silently continuing without sharding.
+                    raise RuntimeError(
+                        "MLX distributed launch requested but initialized a singleton "
+                        "group (size 1). Ensure the installed MLX has distributed "
+                        "support and the launch environment/hostfile is valid, or run "
+                        "without distributed."
+                    )
+                logger.info(
+                    "MLX distributed initialized in worker: rank=%s size=%s mode=%s",
+                    rank,
+                    size,
+                    config.get("mlx_parallel_mode"),
+                )
             _send_response(
                 resp_queue,
                 {"type": "status", "message": "Loading model..."},
@@ -764,6 +864,8 @@ def run_inference_process(
                     if _drain_skip_generate(cmd, resp_queue, drain_event):
                         continue
                     _handle_generate(backend, cmd, resp_queue, cancel_event)
+                elif cmd_type == "share_object":
+                    _handle_share_object(backend, cmd, resp_queue)
                 elif cmd_type == "load":
                     if backend.active_model_name:
                         backend.unload_model(backend.active_model_name)
@@ -977,6 +1079,9 @@ def run_inference_process(
                     continue
                 _handle_generate(backend, cmd, resp_queue, cancel_event)
 
+            elif cmd_type == "share_object":
+                _handle_share_object(backend, cmd, resp_queue)
+
             elif cmd_type == "load":
                 if backend.active_model_name:
                     backend.unload_model(backend.active_model_name)
diff --git a/studio/backend/core/training/trainer.py b/studio/backend/core/training/trainer.py
index 20b2305a5a..96d1b90b16 100644
--- a/studio/backend/core/training/trainer.py
+++ b/studio/backend/core/training/trainer.py
@@ -62,7 +62,6 @@ from loggers import get_logger
 import time
 from pathlib import Path
 from typing import Any, Dict, List, Optional, Callable
-from dataclasses import dataclass
 import pandas as pd
 from datasets import Dataset
 from utils.datasets.cache_safe import load_dataset_cache_safe as load_dataset
@@ -71,7 +70,7 @@ from core.inference.llama_cpp import _hf_offline_if_dns_dead
 from utils.models import is_vision_model, detect_audio_type
 from utils.models.model_config import _env_offline
 from utils.datasets import format_and_template_dataset
-from utils.datasets import MODEL_TO_TEMPLATE_MAPPER, TEMPLATE_TO_RESPONSES_MAPPER
+from utils.datasets.completion_masking import apply_completion_masking
 from utils.datasets.iterable import is_streaming_dataset as detect_streaming_dataset
 from utils.datasets.raw_text import prepare_raw_text_dataset, resolve_column_names
 from utils.paths import (
@@ -86,6 +85,11 @@ from utils.native_path_leases import child_env_without_native_path_secret
 from utils.subprocess_compat import (
     windows_hidden_subprocess_kwargs as _windows_hidden_subprocess_kwargs,
 )
+from .training import (
+    TrainingProgress,
+    create_mlx_trainer_adapter,
+    should_use_mlx_training_backend,
+)
 
 logger = get_logger(__name__)
 
@@ -104,31 +108,16 @@ def _build_report_targets(training_args) -> list[str] | str:
     return report_to or "none"
 
 
-@dataclass
-class TrainingProgress:
-    """Training progress tracking"""
-
-    epoch: float = 0
-    step: int = 0
-    total_steps: int = 0
-    loss: Optional[float] = None
-    learning_rate: Optional[float] = None
-    is_training: bool = False
-    is_completed: bool = False
-    error: Optional[str] = None
-    status_message: str = "Ready to train"  # Current stage
-    elapsed_seconds: Optional[float] = None
-    eta_seconds: Optional[float] = None
-    grad_norm: Optional[float] = None
-    num_tokens: Optional[int] = None
-    eval_loss: Optional[float] = None
-
-
 class UnslothTrainer:
     """
     Unsloth Training Backend
     """
 
+    def __new__(cls, *args, **kwargs):
+        if cls is UnslothTrainer and should_use_mlx_training_backend():
+            return create_mlx_trainer_adapter(*args, **kwargs)
+        return super().__new__(cls)
+
     def __init__(self):
         self.model = None
         self.tokenizer = None
@@ -3466,8 +3455,6 @@ class UnslothTrainer:
 
             # ========== TRAIN ON RESPONSES ONLY ==========
             # Raw-text datasets always train on all tokens.
-            instruction_part = None
-            response_part = None
             is_cpt = training_args.get("is_cpt", False)
             train_on_responses_enabled = (
                 False
@@ -3484,113 +3471,93 @@ class UnslothTrainer:
 
             # DeepSeek OCR handles this internally in its collator, so skip
             # Audio VLM handles label masking in its collator, so skip
+            # Markers auto-detected from the chat template first, manual table
+            # as fallback; gpt-oss stays on its manual markers. See
+            # apply_completion_masking.
             if (
                 train_on_responses_enabled
                 and not self.is_audio_vlm
                 and not self.is_audio
                 and not (is_deepseek_ocr or dataset_final_format == "alpaca")
             ):
-                try:
-                    logger.info("Configuring train on responses only...\n")
+                from unsloth.chat_templates import train_on_responses_only
 
-                    # Template mapping for this model
-                    model_name_lower = self.model_name.lower()
+                logger.info("Configuring train on responses only...\n")
 
-                    if model_name_lower in MODEL_TO_TEMPLATE_MAPPER:
-                        template_name = MODEL_TO_TEMPLATE_MAPPER[model_name_lower]
-                        logger.info(f"Detected template: {template_name}\n")
+                def _notify(level, message):
+                    if level == "warning":
+                        logger.warning(message)
+                    else:
+                        logger.info(f"{message}\n")
 
-                        if template_name in TEMPLATE_TO_RESPONSES_MAPPER:
-                            instruction_part = TEMPLATE_TO_RESPONSES_MAPPER[template_name][
-                                "instruction"
-                            ]
-                            response_part = TEMPLATE_TO_RESPONSES_MAPPER[template_name]["response"]
+                # No try/except: the helper handles detection failures and
+                # double misses itself, so an exception here is a real masking
+                # failure that must fail the run, not silently train on full
+                # sequences.
+                self.trainer, masking_applied = apply_completion_masking(
+                    self.trainer,
+                    self.model_name,
+                    train_on_responses_only,
+                    num_proc = config_args["dataset_num_proc"],
+                    notify = _notify,
+                )
 
-                            logger.info(f"Instruction marker: {instruction_part[:50]}...\n")
-                            logger.info(f"Response marker: {response_part[:50]}...\n")
+                if not masking_applied:
+                    train_on_responses_enabled = False
+
+                if masking_applied:
+                    try:
+                        # ── Safety net: check if all samples were filtered out ──
+                        # train_on_responses_only masks non-response tokens with -100; a
+                        # row becomes all -100 (Unsloth drops it) when the response
+                        # template is not found in the formatted text. Usually a
+                        # dataset/template mismatch (already-formatted data, or 'Train on
+                        # completions' on data that doesn't match the model's chat
+                        # template); only sometimes max_seq_length truncating the response
+                        # away. Skip this len()-based check for streaming.
+                        if detect_streaming_dataset(self.trainer.train_dataset):
+                            logger.info("Skipping post-filter length check for streaming dataset\n")
                         else:
-                            logger.info(
-                                f"No response mapping found for template: {template_name}\n"
+                            filtered_len = len(self.trainer.train_dataset)
+                            original_dataset_obj = (
+                                dataset["dataset"] if isinstance(dataset, dict) else dataset
                             )
-                            train_on_responses_enabled = False
-                    else:
-                        logger.info(f"No template mapping found for model: {self.model_name}\n")
-                        train_on_responses_enabled = False
-
-                except Exception as e:
-                    logger.warning(f"Could not configure train on responses: {e}")
-                    train_on_responses_enabled = False
-
-            # Apply train on responses only if we have valid parts
-            if (
-                train_on_responses_enabled
-                and instruction_part
-                and response_part
-                and not self.is_audio_vlm
-                and not self.is_audio
-                and not (is_deepseek_ocr or dataset_final_format == "alpaca")
-            ):
-                try:
-                    from unsloth.chat_templates import train_on_responses_only
-
-                    self.trainer = train_on_responses_only(
-                        self.trainer,
-                        instruction_part = instruction_part,
-                        response_part = response_part,
-                        num_proc = config_args["dataset_num_proc"],
-                    )
-                    logger.info("Train on responses only configured successfully\n")
-
-                    # ── Safety net: check if all samples were filtered out ──
-                    # train_on_responses_only masks non-response tokens with -100;
-                    # a row becomes all -100 (and Unsloth drops it) when the response
-                    # template is not found in the formatted text. That is usually a
-                    # dataset/template mismatch (already-formatted data, or 'Train on
-                    # completions' applied to data that doesn't match the model's chat
-                    # template), and only sometimes max_seq_length truncating the
-                    # response away. Skip this len()-based check for streaming.
-                    if detect_streaming_dataset(self.trainer.train_dataset):
-                        logger.info("Skipping post-filter length check for streaming dataset\n")
-                    else:
-                        filtered_len = len(self.trainer.train_dataset)
-                        original_dataset_obj = (
-                            dataset["dataset"] if isinstance(dataset, dict) else dataset
-                        )
-                        original_len = len(original_dataset_obj)
-                        dropped = original_len - filtered_len
-                        drop_pct = round(100 * dropped / original_len, 1) if original_len > 0 else 0
-
-                        if filtered_len == 0 or drop_pct > 30:
-                            max_seq = training_args.get("max_seq_length", 2048)
-                            error_msg = (
-                                f"{dropped}/{original_len} samples ({drop_pct}%) were "
-                                f"dropped after applying 'Train on completions': after "
-                                f"masking, those rows had no trainable response tokens "
-                                f"left. The usual cause is that this model's response "
-                                f"template was not found in the formatted samples, so "
-                                f"every token was masked out. That typically means the "
-                                f"dataset is already formatted, or its structure does "
-                                f"not match the model's chat template, so 'Train on "
-                                f"completions' should be turned off for this dataset. "
-                                f"Less commonly, a max_seq_length ({max_seq}) shorter "
-                                f"than the prompt can truncate the response away; only "
-                                f"raise it if your samples are actually longer than that."
+                            original_len = len(original_dataset_obj)
+                            dropped = original_len - filtered_len
+                            drop_pct = (
+                                round(100 * dropped / original_len, 1) if original_len > 0 else 0
                             )
-                            logger.error(error_msg)
-                            self._update_progress(error = error_msg, is_training = False)
-                            return
 
-                        if dropped > 0:
-                            logger.info(
-                                f"⚠️ {dropped}/{original_len} samples "
-                                f"({drop_pct}%) were dropped (all labels "
-                                f"masked). {filtered_len} samples remain.\n"
-                            )
-                        logger.info(f"Post-filter dataset size: {filtered_len} samples\n")
+                            if filtered_len == 0 or drop_pct > 30:
+                                max_seq = training_args.get("max_seq_length", 2048)
+                                error_msg = (
+                                    f"{dropped}/{original_len} samples ({drop_pct}%) were "
+                                    f"dropped after applying 'Train on completions': after "
+                                    f"masking, those rows had no trainable response tokens "
+                                    f"left. The usual cause is that this model's response "
+                                    f"template was not found in the formatted samples, so "
+                                    f"every token was masked out. That typically means the "
+                                    f"dataset is already formatted, or its structure does "
+                                    f"not match the model's chat template, so 'Train on "
+                                    f"completions' should be turned off for this dataset. "
+                                    f"Less commonly, a max_seq_length ({max_seq}) shorter "
+                                    f"than the prompt can truncate the response away; only "
+                                    f"raise it if your samples are actually longer than that."
+                                )
+                                logger.error(error_msg)
+                                self._update_progress(error = error_msg, is_training = False)
+                                return
 
-                except Exception as e:
-                    logger.warning(f"Failed to apply train on responses only: {e}")
-                    train_on_responses_enabled = False
+                            if dropped > 0:
+                                logger.info(
+                                    f"⚠️ {dropped}/{original_len} samples "
+                                    f"({drop_pct}%) were dropped (all labels "
+                                    f"masked). {filtered_len} samples remain.\n"
+                                )
+                            logger.info(f"Post-filter dataset size: {filtered_len} samples\n")
+
+                    except Exception as e:
+                        logger.warning(f"Post-masking dataset size check failed: {e}")
             else:
                 if train_on_responses_enabled and is_deepseek_ocr:
                     logger.info("Train on responses handled by DeepSeek OCR collator\n")
diff --git a/studio/backend/core/training/training.py b/studio/backend/core/training/training.py
index f4233fcf04..2ddda19951 100644
--- a/studio/backend/core/training/training.py
+++ b/studio/backend/core/training/training.py
@@ -14,17 +14,19 @@ import json as _json
 import math
 import multiprocessing as mp
 import os
+import platform
 import queue
 import re
 import shutil
 import threading
 import time
+import traceback
 import structlog
 from datetime import datetime, timezone
 from loggers import get_logger
-from dataclasses import dataclass, field
+from dataclasses import dataclass, field, replace
 from pathlib import Path
-from typing import Optional, Tuple, Any, TYPE_CHECKING
+from typing import Optional, Tuple, Any, Callable, Union, TYPE_CHECKING
 
 if TYPE_CHECKING:
     import matplotlib.pyplot as plt
@@ -98,6 +100,107 @@ def _coerce_optional_nonneg_float(name: str, value):
     return coerced
 
 
+def is_apple_silicon_training_platform() -> bool:
+    return platform.system() == "Darwin" and platform.machine() == "arm64"
+
+
+def is_mlx_training_device(device: Any) -> bool:
+    return (
+        str(device).lower() == "mlx"
+        or str(device).lower().endswith(".mlx")
+        or getattr(device, "name", "").lower() == "mlx"
+    )
+
+
+def should_use_mlx_training_backend(*, device: Optional[Any] = None) -> bool:
+    if device is not None:
+        return is_mlx_training_device(device)
+    return is_apple_silicon_training_platform()
+
+
+def _build_training_worker_config(values: dict[str, Any]) -> dict[str, Any]:
+    """Build the normalized worker config shared by Studio and the CLI adapter."""
+    config = {
+        "model_name": values["model_name"],
+        "project_name": values.get("project_name"),
+        "training_type": values.get("training_type", "LoRA/QLoRA"),
+        "hf_token": values.get("hf_token", ""),
+        "load_in_4bit": values.get("load_in_4bit", True),
+        "max_seq_length": values.get("max_seq_length", 2048),
+        "vision_image_size": values.get("vision_image_size"),
+        "hf_dataset": values.get("hf_dataset", ""),
+        "local_datasets": values.get("local_datasets"),
+        "local_eval_datasets": values.get("local_eval_datasets"),
+        "format_type": values.get("format_type", ""),
+        "subset": values.get("subset"),
+        "train_split": values.get("train_split", "train"),
+        "eval_split": values.get("eval_split"),
+        "eval_steps": values.get("eval_steps", 0.00),
+        "dataset_streaming": values.get("dataset_streaming", False),
+        "dataset_slice_start": values.get("dataset_slice_start"),
+        "dataset_slice_end": values.get("dataset_slice_end"),
+        "custom_format_mapping": values.get("custom_format_mapping"),
+        "is_dataset_image": values.get("is_dataset_image", False),
+        "is_dataset_audio": values.get("is_dataset_audio", False),
+        "is_embedding": values.get("is_embedding", False),
+        "num_epochs": values.get("num_epochs", 3),
+        "learning_rate": values.get("learning_rate", "2e-4"),
+        "embedding_learning_rate": values.get("embedding_learning_rate"),
+        "batch_size": values.get("batch_size", 2),
+        "gradient_accumulation_steps": values.get("gradient_accumulation_steps", 4),
+        "warmup_steps": values.get("warmup_steps"),
+        "warmup_ratio": values.get("warmup_ratio"),
+        "max_steps": values.get("max_steps", 0),
+        "save_steps": values.get("save_steps", 0),
+        "weight_decay": values.get("weight_decay", 0.001),
+        "max_grad_norm": values.get("max_grad_norm", 0.0),
+        "max_grad_value": _coerce_optional_nonneg_float(
+            "max_grad_value", values.get("max_grad_value")
+        ),
+        "max_grad_leaf_norm": _coerce_optional_nonneg_float(
+            "max_grad_leaf_norm", values.get("max_grad_leaf_norm")
+        ),
+        "cast_norm_output_to_input_dtype": _coerce_optional_bool(
+            values.get("cast_norm_output_to_input_dtype"), True
+        ),
+        "random_seed": _coerce_seed(values.get("random_seed")),
+        "packing": values.get("packing", False),
+        "optim": values.get("optim", "adamw_8bit"),
+        "lr_scheduler_type": values.get("lr_scheduler_type", "linear"),
+        "use_lora": values.get("use_lora", True),
+        "lora_r": values.get("lora_r", 16),
+        "lora_alpha": values.get("lora_alpha", 16),
+        "lora_dropout": values.get("lora_dropout", 0.0),
+        "target_modules": values.get("target_modules"),
+        "gradient_checkpointing": values.get("gradient_checkpointing", "unsloth"),
+        "use_rslora": values.get("use_rslora", False),
+        "use_loftq": values.get("use_loftq", False),
+        "train_on_completions": values.get("train_on_completions", False),
+        "finetune_vision_layers": values.get("finetune_vision_layers", True),
+        "finetune_language_layers": values.get("finetune_language_layers", True),
+        "finetune_attention_modules": values.get("finetune_attention_modules", True),
+        "finetune_mlp_modules": values.get("finetune_mlp_modules", True),
+        "enable_wandb": values.get("enable_wandb", False),
+        "wandb_token": values.get("wandb_token"),
+        "wandb_project": values.get("wandb_project", "unsloth-training"),
+        "enable_tensorboard": values.get("enable_tensorboard", False),
+        "tensorboard_dir": values.get("tensorboard_dir", "runs"),
+        "resume_from_checkpoint": values.get("resume_from_checkpoint"),
+        "trust_remote_code": values.get("trust_remote_code", False),
+        "approved_remote_code_fingerprint": values.get("approved_remote_code_fingerprint"),
+        "subject": values.get("subject"),
+        "gpu_ids": values.get("gpu_ids"),
+        "s3_config": values.get("s3_config"),
+        "disable_xet": values.get("disable_xet", False),
+    }
+    for key in ("output_dir", "allow_external_output_dir"):
+        if key in values:
+            config[key] = values.get(key)
+    if config["training_type"] == "Full Finetuning":
+        config["load_in_4bit"] = False
+    return config
+
+
 _HF_TMP_CHECKPOINT_RE = re.compile(r"^tmp-checkpoint-\d+$")
 
 
@@ -133,7 +236,7 @@ def _s3_dataset_name(s3_dataset: Any) -> Optional[str]:
     return f"s3://{bucket}/{prefix}" if prefix else f"s3://{bucket}"
 
 
-def _cleanup_cancelled_checkpoints(output_dir: str | os.PathLike) -> None:
+def _cleanup_cancelled_checkpoints(output_dir: Union[str, os.PathLike]) -> None:
     """Remove only HF Trainer ``tmp-checkpoint-<step>/`` partials after a cancel.
 
     Completed ``checkpoint-<int>/`` dirs survive. Symlinked output_dir / children
@@ -183,7 +286,7 @@ PLOT_HEIGHT = 3.5
 
 @dataclass
 class TrainingProgress:
-    """Mirror of trainer.TrainingProgress so the parent never imports heavy ML modules."""
+    """Shared training progress payload for Studio and backend-aware trainers."""
 
     epoch: float = 0
     step: int = 0
@@ -200,6 +303,423 @@ class TrainingProgress:
     num_tokens: Optional[int] = None
     eval_loss: Optional[float] = None
     peak_memory_gb: Optional[float] = None
+    output_dir: Optional[str] = None
+
+
+class _MLXTrainerAdapter:
+    """Adapts the legacy UnslothTrainer API to the shared Studio MLX worker path."""
+
+    def __init__(self):
+        self.model = None
+        self.tokenizer = None
+        self.trainer = None
+        self.training_thread = None
+        self.training_progress = TrainingProgress()
+        self.progress_callbacks: list[Callable[[TrainingProgress], None]] = []
+        self.is_training = False
+        self.should_stop = False
+        self.save_on_stop = True
+        self.load_in_4bit = True
+        self.output_dir = None
+
+        self.is_cpt = False
+        self.is_vlm = False
+        self.is_audio = False
+        self.is_audio_vlm = False
+        self.model_name = None
+        self.max_seq_length = None
+
+        self._model_config: dict[str, Any] = {}
+        self._peft_config: dict[str, Any] = {}
+        self._dataset_config: dict[str, Any] = {}
+        self._event_queue: Optional[queue.Queue] = None
+        self._stop_queue: Optional[queue.Queue] = None
+        self._pump_thread: Optional[threading.Thread] = None
+        self._lock = threading.Lock()
+
+    def _activate_transformers_for_model(self, model_name: str, hf_token: Optional[str]) -> None:
+        try:
+            from utils.transformers_version import activate_transformers_for_subprocess
+            activate_transformers_for_subprocess(model_name, hf_token)
+        except Exception as exc:
+            logger.warning("MLX trainer adapter Transformers activation failed", error = str(exc))
+
+    def add_progress_callback(self, callback: Callable[[TrainingProgress], None]):
+        self.progress_callbacks.append(callback)
+
+    def _update_progress(self, **kwargs):
+        with self._lock:
+            for key, value in kwargs.items():
+                if hasattr(self.training_progress, key):
+                    setattr(self.training_progress, key, value)
+            progress = self.training_progress
+        for callback in self.progress_callbacks:
+            try:
+                callback(progress)
+            except Exception:
+                pass
+
+    def load_model(
+        self,
+        model_name: str,
+        max_seq_length: int = 2048,
+        load_in_4bit: bool = True,
+        hf_token: Optional[str] = None,
+        is_dataset_image: bool = False,
+        is_dataset_audio: bool = False,
+        trust_remote_code: bool = False,
+        full_finetuning: bool = False,
+        gpu_ids: Optional[list[int]] = None,
+    ) -> bool:
+        self.model_name = model_name
+        self.max_seq_length = max_seq_length
+        self.load_in_4bit = load_in_4bit
+        self._audio_type = None
+        self._activate_transformers_for_model(model_name, hf_token)
+        try:
+            from utils.models import detect_audio_type, is_vision_model
+
+            self._audio_type = detect_audio_type(model_name, hf_token)
+            if self._audio_type == "audio_vlm":
+                self.is_audio = False
+                self.is_audio_vlm = bool(is_dataset_audio)
+                self._audio_type = None
+            else:
+                self.is_audio = self._audio_type is not None
+                self.is_audio_vlm = False
+            vision = is_vision_model(model_name, hf_token = hf_token) if not self.is_audio else False
+            self.is_vlm = not self.is_audio_vlm and vision and bool(is_dataset_image)
+        except Exception as exc:
+            logger.warning("MLX trainer adapter model type detection failed", error = str(exc))
+            self.is_vlm = False
+            self.is_audio = False
+            self.is_audio_vlm = False
+        self.model = object()
+        self.tokenizer = object()
+        self._model_config = {
+            "model_name": model_name,
+            "max_seq_length": max_seq_length,
+            "load_in_4bit": load_in_4bit,
+            "hf_token": hf_token or "",
+            "is_dataset_image": bool(is_dataset_image),
+            "is_dataset_audio": bool(is_dataset_audio),
+            "trust_remote_code": bool(trust_remote_code),
+            "gpu_ids": gpu_ids,
+        }
+        self._update_progress(
+            is_training = False,
+            is_completed = False,
+            error = None,
+            step = 0,
+            loss = 0.0,
+            epoch = 0,
+            status_message = f"Queued MLX model load: {model_name}",
+        )
+        return True
+
+    def prepare_model_for_training(
+        self,
+        use_lora: bool = True,
+        finetune_vision_layers: bool = True,
+        finetune_language_layers: bool = True,
+        finetune_attention_modules: bool = True,
+        finetune_mlp_modules: bool = True,
+        target_modules: Optional[Union[list, str]] = None,
+        lora_r: int = 16,
+        lora_alpha: int = 16,
+        lora_dropout: float = 0.0,
+        use_gradient_checkpointing: Union[str, bool] = "unsloth",
+        use_rslora: bool = False,
+        use_loftq: bool = False,
+    ) -> bool:
+        self._peft_config = {
+            "use_lora": bool(use_lora),
+            "lora_r": lora_r,
+            "lora_alpha": lora_alpha,
+            "lora_dropout": lora_dropout,
+            "target_modules": target_modules,
+            "gradient_checkpointing": use_gradient_checkpointing,
+            "use_rslora": bool(use_rslora),
+            "use_loftq": bool(use_loftq),
+            "finetune_vision_layers": bool(finetune_vision_layers),
+            "finetune_language_layers": bool(finetune_language_layers),
+            "finetune_attention_modules": bool(finetune_attention_modules),
+            "finetune_mlp_modules": bool(finetune_mlp_modules),
+        }
+        self._update_progress(status_message = "Queued MLX training setup")
+        return True
+
+    def load_and_format_dataset(
+        self,
+        dataset_source: Optional[str],
+        format_type: str = "auto",
+        local_datasets: Optional[list[str]] = None,
+        local_eval_datasets: Optional[list[str]] = None,
+        custom_format_mapping: Optional[dict[str, Any]] = None,
+        subset: Optional[str] = None,
+        train_split: str = "train",
+        eval_split: Optional[str] = None,
+        dataset_streaming: bool = False,
+        eval_steps: float = 0.00,
+        dataset_slice_start: Optional[int] = None,
+        dataset_slice_end: Optional[int] = None,
+        is_cpt: bool = False,
+        s3_config: dict = None,
+    ) -> Optional[tuple]:
+        self._dataset_config = {
+            "hf_dataset": dataset_source or "",
+            "local_datasets": local_datasets,
+            "local_eval_datasets": local_eval_datasets,
+            "format_type": format_type or "",
+            "custom_format_mapping": custom_format_mapping,
+            "subset": subset,
+            "train_split": train_split or "train",
+            "eval_split": eval_split,
+            "dataset_streaming": bool(dataset_streaming),
+            "eval_steps": eval_steps or 0.0,
+            "dataset_slice_start": dataset_slice_start,
+            "dataset_slice_end": dataset_slice_end,
+            "s3_config": s3_config,
+        }
+        self.is_cpt = bool(is_cpt)
+        self._update_progress(status_message = "Queued MLX dataset load")
+        return ({"dataset": [], "final_format": "deferred_mlx_cli", "success": True}, None)
+
+    def start_training(
+        self,
+        dataset = None,
+        eval_dataset = None,
+        **training_args,
+    ) -> bool:
+        if self.is_training and self.training_thread and self.training_thread.is_alive():
+            return False
+        if self._pump_thread and self._pump_thread.is_alive():
+            self._pump_thread.join(timeout = 2.0)
+            if self._pump_thread.is_alive():
+                self._update_progress(error = "Previous training event pump is still finalizing")
+                return False
+        if not self._model_config:
+            self._update_progress(error = "Model not loaded")
+            return False
+        if not self._dataset_config:
+            self._update_progress(error = "Dataset not loaded")
+            return False
+        if self.is_cpt:
+            self._update_progress(
+                error = "Continued Pretraining is not supported for MLX training yet.",
+                is_training = False,
+                is_completed = False,
+            )
+            return False
+
+        config = self._build_worker_config(training_args)
+        event_queue = queue.Queue()
+        stop_queue = queue.Queue()
+        self._event_queue = event_queue
+        self._stop_queue = stop_queue
+        self.should_stop = False
+        self.is_training = True
+        self.training_progress = TrainingProgress(
+            is_training = True,
+            status_message = "Initializing MLX training...",
+        )
+
+        self.training_thread = threading.Thread(
+            target = self._run_training_thread,
+            args = (config, event_queue, stop_queue),
+            daemon = True,
+        )
+        self._pump_thread = threading.Thread(
+            target = self._pump_events,
+            args = (event_queue, self.training_thread),
+            daemon = True,
+        )
+        self.training_thread.start()
+        self._pump_thread.start()
+        return True
+
+    def _build_worker_config(self, training_args: dict[str, Any]) -> dict[str, Any]:
+        peft = {
+            "use_lora": True,
+            "lora_r": 16,
+            "lora_alpha": 16,
+            "lora_dropout": 0.0,
+            "target_modules": None,
+            "gradient_checkpointing": "unsloth",
+            "use_rslora": False,
+            "use_loftq": False,
+            "finetune_vision_layers": True,
+            "finetune_language_layers": True,
+            "finetune_attention_modules": True,
+            "finetune_mlp_modules": True,
+            **self._peft_config,
+        }
+        output_dir = training_args.get("output_dir")
+        if output_dir:
+            output_dir = os.path.abspath(os.path.expanduser(str(output_dir)))
+        values = {
+            **self._model_config,
+            **self._dataset_config,
+            **training_args,
+            "training_type": (
+                "Continued Pretraining"
+                if self.is_cpt
+                else "LoRA/QLoRA"
+                if peft["use_lora"]
+                else "Full Finetuning"
+            ),
+            **peft,
+            "output_dir": output_dir,
+            "allow_external_output_dir": bool(output_dir),
+        }
+        config = _build_training_worker_config(values)
+        config["resolved_gpu_ids"] = None
+        config["gpu_selection"] = None
+        return config
+
+    def _run_training_thread(
+        self, config: dict[str, Any], event_queue: queue.Queue, stop_queue: queue.Queue
+    ):
+        try:
+            self._run_mlx_worker(config, event_queue, stop_queue)
+        except Exception as exc:
+            if event_queue is not None:
+                event_queue.put(
+                    {
+                        "type": "error",
+                        "error": str(exc),
+                        "stack": traceback.format_exc(limit = 20),
+                        "ts": time.time(),
+                    }
+                )
+
+    def _run_mlx_worker(
+        self, config: dict[str, Any], event_queue: queue.Queue, stop_queue: queue.Queue
+    ):
+        from .worker import run_mlx_training_process
+        run_mlx_training_process(
+            event_queue = event_queue,
+            stop_queue = stop_queue,
+            config = config,
+        )
+
+    def _pump_events(self, event_queue: queue.Queue, training_thread: threading.Thread):
+        while True:
+            event = None
+            try:
+                event = event_queue.get(timeout = 0.25)
+            except queue.Empty:
+                pass
+            if event is not None:
+                self._handle_event(event)
+                continue
+            if not training_thread.is_alive():
+                self._drain_events(event_queue)
+                with self._lock:
+                    if self.training_progress.is_training:
+                        self.training_progress.is_training = False
+                        if self.should_stop:
+                            self.training_progress.status_message = "Training stopped."
+                        elif (
+                            not self.training_progress.error
+                            and not self.training_progress.is_completed
+                        ):
+                            self.training_progress.error = "Training process exited unexpectedly"
+                    self.is_training = False
+                    self._event_queue = None
+                    self._stop_queue = None
+                return
+
+    def _drain_events(self, event_queue: Optional[queue.Queue] = None):
+        event_queue = event_queue or self._event_queue
+        if event_queue is None:
+            return
+        while True:
+            try:
+                self._handle_event(event_queue.get_nowait())
+            except queue.Empty:
+                return
+
+    def _handle_event(self, event: dict[str, Any]):
+        etype = event.get("type")
+        if etype == "status":
+            self._update_progress(
+                status_message = event.get("status_message") or event.get("message") or ""
+            )
+            return
+        if etype == "progress":
+            self._update_progress(
+                step = event.get("step", self.training_progress.step),
+                epoch = event.get("epoch", self.training_progress.epoch),
+                loss = event.get("loss", self.training_progress.loss),
+                learning_rate = event.get("learning_rate", self.training_progress.learning_rate),
+                total_steps = event.get("total_steps", self.training_progress.total_steps),
+                elapsed_seconds = event.get(
+                    "elapsed_seconds",
+                    self.training_progress.elapsed_seconds,
+                ),
+                eta_seconds = event.get("eta_seconds", self.training_progress.eta_seconds),
+                grad_norm = event.get("grad_norm", self.training_progress.grad_norm),
+                num_tokens = event.get("num_tokens", self.training_progress.num_tokens),
+                eval_loss = event.get("eval_loss", self.training_progress.eval_loss),
+                peak_memory_gb = event.get("peak_memory_gb", self.training_progress.peak_memory_gb),
+            )
+            return
+        if etype == "complete":
+            status_message = event.get("status_message") or "Training completed"
+            output_dir = event.get("output_dir")
+            was_cancelled = self.should_stop or status_message.strip().lower() in {
+                "training cancelled",
+                "training stopped",
+            }
+            self.output_dir = output_dir
+            self._update_progress(
+                is_training = False,
+                is_completed = not was_cancelled,
+                error = None,
+                status_message = status_message,
+                output_dir = output_dir,
+            )
+            self.is_training = False
+            return
+        if etype == "error":
+            self._update_progress(
+                is_training = False,
+                is_completed = False,
+                error = event.get("error") or event.get("message") or "Training failed",
+            )
+            self.is_training = False
+            return
+
+    def stop_training(self, save: bool = True):
+        self.should_stop = True
+        self.save_on_stop = bool(save)
+        if self._stop_queue is not None:
+            self._stop_queue.put({"type": "stop", "save": save})
+        status_message = (
+            "Stopping training and saving checkpoint..." if save else "Cancelling training..."
+        )
+        self._update_progress(status_message = status_message)
+        return True
+
+    def get_training_progress(self) -> TrainingProgress:
+        pump_thread = self._pump_thread
+        training_thread = self.training_thread
+        if (
+            pump_thread is not None
+            and pump_thread.is_alive()
+            and (training_thread is None or not training_thread.is_alive())
+            and threading.current_thread() is not pump_thread
+        ):
+            pump_thread.join(timeout = 5.0)
+        if pump_thread is None or not pump_thread.is_alive():
+            self._drain_events()
+        with self._lock:
+            return replace(self.training_progress)
+
+
+def create_mlx_trainer_adapter(*args, **kwargs):
+    return _MLXTrainerAdapter(*args, **kwargs)
 
 
 class TrainingBackend:
@@ -296,86 +816,7 @@ class TrainingBackend:
         # treat this fresh setup as a recoverable death.
         self._pump_running = False
 
-        # Build config dict for the subprocess
-        config = {
-            "model_name": kwargs["model_name"],
-            "project_name": kwargs.get("project_name"),
-            "training_type": kwargs.get("training_type", "LoRA/QLoRA"),
-            "hf_token": kwargs.get("hf_token", ""),
-            "load_in_4bit": kwargs.get("load_in_4bit", True),
-            "max_seq_length": kwargs.get("max_seq_length", 2048),
-            "vision_image_size": kwargs.get("vision_image_size"),
-            "hf_dataset": kwargs.get("hf_dataset", ""),
-            "local_datasets": kwargs.get("local_datasets"),
-            "local_eval_datasets": kwargs.get("local_eval_datasets"),
-            "format_type": kwargs.get("format_type", ""),
-            "subset": kwargs.get("subset"),
-            "train_split": kwargs.get("train_split", "train"),
-            "eval_split": kwargs.get("eval_split"),
-            "eval_steps": kwargs.get("eval_steps", 0.00),
-            "dataset_streaming": kwargs.get("dataset_streaming", False),
-            "dataset_slice_start": kwargs.get("dataset_slice_start"),
-            "dataset_slice_end": kwargs.get("dataset_slice_end"),
-            "custom_format_mapping": kwargs.get("custom_format_mapping"),
-            "is_dataset_image": kwargs.get("is_dataset_image", False),
-            "is_dataset_audio": kwargs.get("is_dataset_audio", False),
-            "is_embedding": kwargs.get("is_embedding", False),
-            "num_epochs": kwargs.get("num_epochs", 3),
-            "learning_rate": kwargs.get("learning_rate", "2e-4"),
-            "embedding_learning_rate": kwargs.get("embedding_learning_rate"),
-            "batch_size": kwargs.get("batch_size", 2),
-            "gradient_accumulation_steps": kwargs.get("gradient_accumulation_steps", 4),
-            "warmup_steps": kwargs.get("warmup_steps"),
-            "warmup_ratio": kwargs.get("warmup_ratio"),
-            "max_steps": kwargs.get("max_steps", 0),
-            "save_steps": kwargs.get("save_steps", 0),
-            "weight_decay": kwargs.get("weight_decay", 0.001),
-            "max_grad_norm": kwargs.get("max_grad_norm", 0.0),
-            "max_grad_value": _coerce_optional_nonneg_float(
-                "max_grad_value", kwargs.get("max_grad_value")
-            ),
-            "max_grad_leaf_norm": _coerce_optional_nonneg_float(
-                "max_grad_leaf_norm", kwargs.get("max_grad_leaf_norm")
-            ),
-            "cast_norm_output_to_input_dtype": _coerce_optional_bool(
-                kwargs.get("cast_norm_output_to_input_dtype"), True
-            ),
-            # MLX/CUDA/embedding workers need an int (transformers.set_seed(None) raises).
-            "random_seed": _coerce_seed(kwargs.get("random_seed")),
-            "packing": kwargs.get("packing", False),
-            "optim": kwargs.get("optim", "adamw_8bit"),
-            "lr_scheduler_type": kwargs.get("lr_scheduler_type", "linear"),
-            "use_lora": kwargs.get("use_lora", True),
-            "lora_r": kwargs.get("lora_r", 16),
-            "lora_alpha": kwargs.get("lora_alpha", 16),
-            "lora_dropout": kwargs.get("lora_dropout", 0.0),
-            "target_modules": kwargs.get("target_modules"),
-            "gradient_checkpointing": kwargs.get("gradient_checkpointing", "unsloth"),
-            "use_rslora": kwargs.get("use_rslora", False),
-            "use_loftq": kwargs.get("use_loftq", False),
-            "train_on_completions": kwargs.get("train_on_completions", False),
-            "finetune_vision_layers": kwargs.get("finetune_vision_layers", True),
-            "finetune_language_layers": kwargs.get("finetune_language_layers", True),
-            "finetune_attention_modules": kwargs.get("finetune_attention_modules", True),
-            "finetune_mlp_modules": kwargs.get("finetune_mlp_modules", True),
-            "enable_wandb": kwargs.get("enable_wandb", False),
-            "wandb_token": kwargs.get("wandb_token"),
-            "wandb_project": kwargs.get("wandb_project", "unsloth-training"),
-            "enable_tensorboard": kwargs.get("enable_tensorboard", False),
-            "tensorboard_dir": kwargs.get("tensorboard_dir", "runs"),
-            "resume_from_checkpoint": kwargs.get("resume_from_checkpoint"),
-            "trust_remote_code": kwargs.get("trust_remote_code", False),
-            "approved_remote_code_fingerprint": kwargs.get("approved_remote_code_fingerprint"),
-            "subject": kwargs.get("subject"),
-            "gpu_ids": kwargs.get("gpu_ids"),
-            "s3_config": kwargs.get("s3_config"),
-            # Flipped to True only by the HTTP-fallback respawn after a stall.
-            "disable_xet": kwargs.get("disable_xet", False),
-        }
-
-        # Full finetuning always runs in 16-bit; LoRA/QLoRA/CPT keep the request.
-        if config["training_type"] == "Full Finetuning":
-            config["load_in_4bit"] = False
+        config = _build_training_worker_config(kwargs)
 
         # Split GPU validation from placement around the VRAM hook:
         #   * Explicit gpu_ids are validated here (raises -> the route returns 400
@@ -401,7 +842,7 @@ class TrainingBackend:
         )
 
         defer_auto_selection = False
-        if _hw.DEVICE == _hw.DeviceType.MLX:
+        if should_use_mlx_training_backend(device = _hw.DEVICE):
             config["resolved_gpu_ids"] = None
             config["gpu_selection"] = None
         elif gpu_ids:
@@ -1022,17 +1463,22 @@ class TrainingBackend:
                 self._progress.is_training = True
 
             elif etype == "complete":
-                self._progress.is_training = False
-                self._progress.is_completed = True
-                self._output_dir = event.get("output_dir")
                 msg = event.get("status_message", "Training completed")
+                stopped = self._should_stop or msg.strip().lower() in {
+                    "training cancelled",
+                    "training stopped",
+                }
+                self._progress.is_training = False
+                self._progress.is_completed = not stopped
+                self._output_dir = event.get("output_dir")
+                self._progress.output_dir = self._output_dir
                 self._progress.status_message = msg
                 if not self._db_run_created and self.current_job_id and self._db_config:
                     db_action = "create_and_finalize"
                 else:
                     db_action = "finalize"
                 db_action_kwargs = {
-                    "status": "stopped" if self._should_stop else "completed",
+                    "status": "stopped" if stopped else "completed",
                     "output_dir": self._output_dir,
                 }
 
diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index 17dc1299ca..5fb8fc2bb6 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -1309,14 +1309,18 @@ def _normalize_mlx_studio_scheduler(value):
 
 
 def _resolve_mlx_local_dataset_files(file_paths: list) -> list[str]:
-    """Resolve Studio local dataset uploads without importing the GPU trainer."""
+    """Resolve CLI paths and Studio local dataset uploads without importing the GPU trainer."""
     from utils.paths import resolve_dataset_path
 
     all_files: list[str] = []
     for dataset_file in file_paths or []:
-        file_path = (
-            dataset_file if os.path.isabs(dataset_file) else str(resolve_dataset_path(dataset_file))
-        )
+        dataset_path = Path(os.path.expanduser(str(dataset_file)))
+        if dataset_path.is_absolute():
+            file_path = str(dataset_path)
+        elif dataset_path.exists():
+            file_path = str(dataset_path.resolve())
+        else:
+            file_path = str(resolve_dataset_path(str(dataset_file)))
         file_path_obj = Path(file_path)
 
         if file_path_obj.is_dir():
@@ -1355,6 +1359,58 @@ def _mlx_local_dataset_loader_for_files(files: list[str]) -> str:
     raise ValueError(f"Unsupported dataset format: {files[0]}")
 
 
+_MLX_WORKER_COMPLETE = "_mlx_worker_complete"
+
+
+def _start_mlx_stop_poller(stop_queue):
+    import queue as _queue
+    import threading
+
+    stop_save = [True]
+    stop_requested = [False]
+    trainer_ref = [None]
+
+    def is_stop_requested():
+        return stop_requested[0]
+
+    def poll_stop():
+        while True:
+            try:
+                msg = stop_queue.get(timeout = 0.25)
+                if msg and msg.get("type") == _MLX_WORKER_COMPLETE:
+                    return
+                if msg and msg.get("type") == "stop":
+                    stop_save[0] = msg.get("save", True)
+                    stop_requested[0] = True
+                    trainer = trainer_ref[0]
+                    if trainer is not None:
+                        trainer.stop_requested = True
+                    return
+            except _queue.Empty:
+                continue
+            except (EOFError, OSError):
+                return
+
+    stop_thread = threading.Thread(target = poll_stop, daemon = True)
+    stop_thread.start()
+    return stop_save, stop_requested, trainer_ref, is_stop_requested, stop_thread
+
+
+def _resolve_mlx_output_dir(config, model_name):
+    from utils.paths import resolve_output_dir, default_run_dir_name
+
+    output_dir = config.get("output_dir", "")
+    if not output_dir:
+        output_dir = f"{default_run_dir_name(model_name)}_{int(time.time())}"
+        return str(resolve_output_dir(output_dir))
+    if config.get("allow_external_output_dir"):
+        output_path = Path(output_dir).expanduser()
+        if not output_path.is_absolute():
+            output_path = Path.cwd() / output_path
+        return str(output_path.resolve())
+    return str(resolve_output_dir(output_dir))
+
+
 def _run_mlx_training(event_queue, stop_queue, config):
     """Self-contained MLX training path for Apple Silicon.
 
@@ -1363,8 +1419,6 @@ def _run_mlx_training(event_queue, stop_queue, config):
     """
     import time
     import math
-    import threading
-    import queue as _queue
     from pathlib import Path
 
     def _send(event_type, **kwargs):
@@ -1374,31 +1428,9 @@ def _run_mlx_training(event_queue, stop_queue, config):
                 kwargs["message"] = sm
         event_queue.put({"type": event_type, "ts": time.time(), **kwargs})
 
-    _stop_save = [True]
-    _stop_requested = [False]
-    _trainer_ref = [None]
-
-    def _is_stop_requested():
-        return _stop_requested[0]
-
-    def _poll_stop():
-        while True:
-            try:
-                msg = stop_queue.get(timeout = 1.0)
-                if msg and msg.get("type") == "stop":
-                    _stop_save[0] = msg.get("save", True)
-                    _stop_requested[0] = True
-                    trainer = _trainer_ref[0]
-                    if trainer is not None:
-                        trainer.stop_requested = True
-                    return
-            except _queue.Empty:
-                continue
-            except (EOFError, OSError):
-                return
-
-    stop_thread = threading.Thread(target = _poll_stop, daemon = True)
-    stop_thread.start()
+    _stop_save, _stop_requested, _trainer_ref, _is_stop_requested, _stop_thread = (
+        _start_mlx_stop_poller(stop_queue)
+    )
 
     _send("status", status_message = "Loading MLX libraries...")
 
@@ -1699,6 +1731,7 @@ def _run_mlx_training(event_queue, stop_queue, config):
     # sharegpt+images) and text (alpaca/sharegpt/chatml → "text" column).
     format_type = config.get("format_type", "")
     custom_format_mapping = config.get("custom_format_mapping")
+    dataset_final_format = ""
     try:
         from utils.datasets import format_and_template_dataset
         def _fmt_progress(status_message = "", **_kw):
@@ -1764,6 +1797,7 @@ def _run_mlx_training(event_queue, stop_queue, config):
             )
             if info.get("success", True):
                 dataset = info.get("dataset", dataset)
+            dataset_final_format = str(info.get("final_format", "") or "").lower()
             if eval_dataset is not None:
                 ev = format_and_template_dataset(
                     eval_dataset,
@@ -1804,21 +1838,14 @@ def _run_mlx_training(event_queue, stop_queue, config):
 
     # ── 5. Build output dir ──
     # Resolve to ~/.unsloth/studio/outputs/ so the export page finds it
-    from utils.paths import resolve_output_dir, ensure_dir
+    from utils.paths import ensure_dir
 
-    output_dir = config.get("output_dir", "")
-    if not output_dir:
-        output_dir = build_default_output_dir_name(
-            model_name,
-            config.get("project_name"),
-        )
-    output_dir = str(resolve_output_dir(output_dir))
+    output_dir = _resolve_mlx_output_dir(config, model_name)
     ensure_dir(Path(output_dir))
 
     # ── 6. Create trainer ──
     eval_steps_val = config.get("eval_steps", 0) or 0
     if isinstance(eval_steps_val, float) and 0 < eval_steps_val < 1:
-        # Studio sometimes sends fraction-of-total-steps
         eval_steps_val = max(1, int(eval_steps_val * max_steps))
     else:
         eval_steps_val = int(eval_steps_val)
@@ -1869,6 +1896,9 @@ def _run_mlx_training(event_queue, stop_queue, config):
         eval_steps = eval_steps_val,
     )
 
+    # Also gates the masking skip below, so defined outside the feature-detect block.
+    raw_text_mode = training_type == "Continued Pretraining" or format_type == "raw"
+
     # Feature-detect optional fields so this PR works without the paired zoo bump.
     _supported_fields = getattr(MLXTrainingConfig, "__dataclass_fields__", {})
     if "cast_norm_output_to_input_dtype" in _supported_fields:
@@ -1882,7 +1912,6 @@ def _run_mlx_training(event_queue, stop_queue, config):
     if "max_grad_leaf_norm" in _supported_fields:
         mlx_config_kwargs["max_grad_leaf_norm"] = max_grad_leaf_norm
     if "append_eos" in _supported_fields:
-        raw_text_mode = training_type == "Continued Pretraining" or format_type == "raw"
         # Studio SFT formatting owns rendered examples; raw/CPT text still
         # needs MLX to append EOS like the CUDA raw-text path.
         mlx_config_kwargs["append_eos"] = bool(raw_text_mode)
@@ -1903,29 +1932,27 @@ def _run_mlx_training(event_queue, stop_queue, config):
         _send("eval_configured")
 
     # ── 7. Apply train_on_responses_only if requested ──
-    if config.get("train_on_completions", False):
+    # Auto-detect markers from the chat template first, manual table as
+    # fallback. Mirror the CUDA skips: raw/CPT text has no chat turns and
+    # Alpaca-rendered text lacks the chat markers. Also check the resolved
+    # format, since format_type="auto" can land on alpaca or raw text.
+    if (
+        config.get("train_on_completions", False)
+        and not raw_text_mode
+        and format_type != "alpaca"
+        and dataset_final_format not in ("alpaca", "raw_text")
+    ):
         _send("status", status_message = "Configuring response-only training...")
-        try:
-            from utils.datasets import (
-                MODEL_TO_TEMPLATE_MAPPER,
-                TEMPLATE_TO_RESPONSES_MAPPER,
-            )
-
-            template_name = MODEL_TO_TEMPLATE_MAPPER.get(model_name.lower())
-            markers = TEMPLATE_TO_RESPONSES_MAPPER.get(template_name) if template_name else None
-            if markers:
-                trainer = train_on_responses_only(
-                    trainer,
-                    instruction_part = markers["instruction"],
-                    response_part = markers["response"],
-                )
-            else:
-                _send(
-                    "status",
-                    status_message = f"train_on_completions skipped (no template for {model_name})",
-                )
-        except Exception as e:
-            _send("status", status_message = f"train_on_completions failed: {e}")
+        # No catch: the helper handles detection failures and double misses, so
+        # an exception here is a real masking failure that must fail the run,
+        # not silently train on full sequences.
+        from utils.datasets.completion_masking import apply_completion_masking
+        trainer, _masking_applied = apply_completion_masking(
+            trainer,
+            model_name,
+            train_on_responses_only,
+            notify = lambda level, message: _send("status", status_message = message),
+        )
 
     # ── 8. Setup wandb / tensorboard ──
     wandb_run = None
@@ -2043,12 +2070,27 @@ def _run_mlx_training(event_queue, stop_queue, config):
     # ── 11. Run training ──
     gc.collect()
     mx.synchronize()
-    trainer.train(resume_from_checkpoint = resume_from_checkpoint)
+    _save_model = trainer.save_model
+
+    def _skip_internal_final_save(*args, **kwargs):
+        raise ValueError("worker owns final save")
+
+    trainer.save_model = _skip_internal_final_save
+    try:
+        trainer.train(resume_from_checkpoint = resume_from_checkpoint)
+    finally:
+        trainer.save_model = _save_model
 
     # ── 12. Save and finalize ──
-    if trainer.stop_requested and not _stop_save[0]:
-        # User clicked "Cancel" (save=False) — skip saving
-        _send("complete", output_dir = None, status_message = "Training cancelled")
+    if trainer.stop_requested:
+        if not _stop_save[0]:
+            # Cancel (save=False): skip saving.
+            _send("complete", output_dir = None, status_message = "Training cancelled")
+        else:
+            _send("status", status_message = "Saving stopped model...")
+            mx.synchronize()
+            trainer.save_model(output_dir)
+            _send("complete", output_dir = output_dir, status_message = "Training stopped")
     else:
         _send("status", status_message = "Saving model...")
         mx.synchronize()
@@ -2067,6 +2109,79 @@ def _run_mlx_training(event_queue, stop_queue, config):
             pass
 
 
+def _is_current_process_apple_silicon() -> bool:
+    import platform
+    return platform.system() == "Darwin" and platform.machine() == "arm64"
+
+
+def run_mlx_training_process(
+    *,
+    event_queue: Any,
+    stop_queue: Any,
+    config: dict,
+    transformers_activated: bool = False,
+) -> None:
+    """MLX worker entrypoint shared by Studio subprocesses and the CLI adapter."""
+    model_name = config["model_name"]
+
+    backend_path = str(Path(__file__).resolve().parent.parent.parent)
+    if backend_path not in sys.path:
+        sys.path.insert(0, backend_path)
+
+    from utils.hf_xet_fallback import child_should_disable_xet
+
+    if child_should_disable_xet(config):
+        os.environ["HF_HUB_DISABLE_XET"] = "1"
+        os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "0"
+
+    if not transformers_activated:
+        # Must precede detect_hardware(): its MLX stack check imports mlx_lm, hence transformers.
+        _activate_transformers_version_or_warn(model_name, config.get("hf_token") or None)
+
+    from utils.hardware import hardware as _hw
+
+    _hw.detect_hardware()
+    if _hw.DEVICE != _hw.DeviceType.MLX:
+        event_queue.put(
+            {
+                "type": "error",
+                "error": "MLX training requires Apple Silicon with the MLX backend available.",
+                "stack": "",
+                "ts": time.time(),
+            }
+        )
+        return
+
+    if config.get("is_dataset_audio"):
+        event_queue.put(
+            {
+                "type": "error",
+                "error": "Audio dataset training is not yet supported on Apple Silicon.",
+                "stack": "",
+                "ts": time.time(),
+            }
+        )
+        return
+
+    try:
+        try:
+            _run_mlx_training(event_queue, stop_queue, config)
+        finally:
+            try:
+                stop_queue.put({"type": _MLX_WORKER_COMPLETE})
+            except (EOFError, OSError, ValueError):
+                pass
+    except Exception as exc:
+        event_queue.put(
+            {
+                "type": "error",
+                "error": str(exc),
+                "stack": traceback.format_exc(limit = 20),
+                "ts": time.time(),
+            }
+        )
+
+
 def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) -> None:
     """Subprocess entrypoint. Fresh Python — no stale module state.
 
@@ -2141,36 +2256,26 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) ->
     if backend_path not in sys.path:
         sys.path.insert(0, backend_path)
 
+    from .training import is_apple_silicon_training_platform, should_use_mlx_training_backend
+
+    mlx_backend_requested = is_apple_silicon_training_platform()
+
+    mlx_transformers_activated = False
+    if mlx_backend_requested and _is_current_process_apple_silicon():
+        # Must precede detect_hardware(): its MLX stack check imports mlx_lm, hence transformers.
+        _activate_transformers_version_or_warn(model_name, config.get("hf_token") or None)
+        mlx_transformers_activated = True
+
     from utils.hardware import hardware as _hw
 
     _hw.detect_hardware()
-    if _hw.DEVICE == _hw.DeviceType.MLX:
-        if config.get("is_dataset_audio"):
-            event_queue.put(
-                {
-                    "type": "error",
-                    "error": "Audio dataset training is not yet supported on Apple Silicon.",
-                    "stack": "",
-                    "ts": time.time(),
-                }
-            )
-            return
-        # Activate correct transformers version (Gemma-4 needs a 5.x sidecar, etc.)
-        # Must happen before any transformers/mlx-lm imports in _run_mlx_training.
-        # Non-fatal: fall through with whatever version is installed, but log
-        # the failure instead of swallowing it (issue #6103).
-        _activate_transformers_version_or_warn(model_name, config.get("hf_token") or None)
-        try:
-            _run_mlx_training(event_queue, stop_queue, config)
-        except Exception as exc:
-            event_queue.put(
-                {
-                    "type": "error",
-                    "error": str(exc),
-                    "stack": traceback.format_exc(limit = 20),
-                    "ts": time.time(),
-                }
-            )
+    if mlx_backend_requested or should_use_mlx_training_backend(device = _hw.DEVICE):
+        run_mlx_training_process(
+            event_queue = event_queue,
+            stop_queue = stop_queue,
+            config = config,
+            transformers_activated = mlx_transformers_activated,
+        )
         return
 
     # ── 1. Activate correct transformers version BEFORE any ML imports ──
@@ -2693,7 +2798,8 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) ->
         if backend_path not in sys.path:
             sys.path.insert(0, backend_path)
 
-        from core.training.trainer import UnslothTrainer, TrainingProgress
+        from core.training.training import TrainingProgress
+        from core.training.trainer import UnslothTrainer
         from utils.paths import (
             ensure_dir,
             resolve_output_dir,
diff --git a/studio/backend/hub/tests/test_model_services.py b/studio/backend/hub/tests/test_model_services.py
index f05d8359ec..44701c0b64 100644
--- a/studio/backend/hub/tests/test_model_services.py
+++ b/studio/backend/hub/tests/test_model_services.py
@@ -105,6 +105,10 @@ def test_repo_id_validation_accepts_hf_repo_id_contract(repo_id):
     assert paths.is_valid_repo_id(repo_id)
 
 
+def test_repo_id_validation_accepts_max_length_namespaced_repo():
+    assert paths.is_valid_repo_id(f"{'a' * 96}/{'b' * 96}")
+
+
 @pytest.mark.parametrize(
     "repo_id",
     [
@@ -121,6 +125,48 @@ def test_repo_id_validation_rejects_unsafe_or_invalid_ids(repo_id):
     assert not paths.is_valid_repo_id(repo_id)
 
 
+def test_download_state_preserves_readable_keys_when_safe(monkeypatch, tmp_path):
+    monkeypatch.setattr(state_dir, "cache_root", lambda: tmp_path)
+
+    path = state_dir.marker_path("model", "Owner/Repo", "Q4_K_M")
+
+    assert path is not None
+    assert path.name == "models--owner--repo--variant--q4_k_m.json"
+
+
+@pytest.mark.parametrize("variant", ["bad variant with spaces", "q" * 64])
+def test_download_state_bounds_long_repo_variant_filenames(monkeypatch, tmp_path, variant):
+    monkeypatch.setattr(state_dir, "cache_root", lambda: tmp_path)
+    repo_id = f"{'a' * 96}/{'b' * 96}"
+
+    assert paths.is_valid_repo_id(repo_id)
+    assert download_manifest.write_cancel_marker("model", repo_id, variant, "http")
+    assert download_manifest.write_manifest(
+        "model",
+        repo_id,
+        variant,
+        [download_manifest.ExpectedFile(path = "model.gguf", size = 1)],
+        "http",
+    )
+
+    marker_path = state_dir.marker_path("model", repo_id, variant)
+    manifest_path = state_dir.manifest_path("model", repo_id, variant)
+
+    assert marker_path is not None
+    assert manifest_path is not None
+    assert "--sha256-" in marker_path.name
+    assert len(marker_path.name.encode("utf-8")) <= 255
+    assert len(f".{marker_path.name}.tmp-00000000".encode("utf-8")) <= 255
+    assert download_manifest.has_cancel_marker("model", repo_id, variant)
+    assert download_manifest.read_manifest("model", repo_id, variant) is not None
+    assert list(download_manifest.iter_variant_markers("model", repo_id)) == [
+        (variant, marker_path)
+    ]
+    assert list(download_manifest.iter_variant_manifests("model", repo_id)) == [
+        (variant, manifest_path)
+    ]
+
+
 class _RecordingLogger:
     def __init__(self):
         self.warnings = []
diff --git a/studio/backend/hub/utils/gguf_plan.py b/studio/backend/hub/utils/gguf_plan.py
index 2abdb0fb79..18daa4f84e 100644
--- a/studio/backend/hub/utils/gguf_plan.py
+++ b/studio/backend/hub/utils/gguf_plan.py
@@ -102,16 +102,17 @@ def preferred_mmproj_sibling(siblings: Sequence) -> Optional[object]:
 def preferred_mtp_sibling(siblings: Sequence) -> Optional[object]:
     """The separate MTP drafter to fetch with every variant: the repo-root
     ``mtp-*.gguf`` copy unsloth ships for llama.cpp ``-hf`` auto-discovery
-    (Gemma 4). Same pick as the loader's drafter resolution (``mtp-`` basename
-    prefix, first in sort order) so download and load resolve the same file;
-    the higher-precision ``MTP/`` subdir copies are for explicit selection and
-    are not auto-fetched. None for repos with the head baked into the main
-    GGUF (Qwen)."""
+    (Gemma 4). Same pick as the loader's drafter resolution (root-level
+    ``mtp-`` prefix, first in sort order) so download and load resolve the same
+    file; the higher-precision ``MTP/`` subdir copies are for explicit
+    selection and are not auto-fetched. None for repos with the head baked into
+    the main GGUF (Qwen)."""
+    # Root-level only: the MTP/ subdir copies now share the mtp- prefix too.
     candidates = sorted(
         (
             s
             for s in siblings
-            if (name := _gguf_rfilename(s)) and name.lower().rsplit("/", 1)[-1].startswith("mtp-")
+            if (name := _gguf_rfilename(s)) and "/" not in name and name.lower().startswith("mtp-")
         ),
         key = lambda s: getattr(s, "rfilename"),
     )
diff --git a/studio/backend/hub/utils/paths.py b/studio/backend/hub/utils/paths.py
index afcb0b41dc..5435202565 100644
--- a/studio/backend/hub/utils/paths.py
+++ b/studio/backend/hub/utils/paths.py
@@ -181,15 +181,20 @@ def is_valid_repo_id(repo_id: str) -> bool:
     """Validate Hugging Face ``repo_name`` or ``namespace/repo_name`` IDs."""
     if not repo_id or repo_id != repo_id.strip():
         return False
-    if len(repo_id) > _MAX_REPO_ID_LENGTH or repo_id.endswith(".git"):
+    if repo_id.endswith(".git"):
         return False
     if "--" in repo_id or ".." in repo_id:
         return False
     segments = repo_id.split("/")
     if len(segments) not in (1, 2):
         return False
+    # Match huggingface_hub.validate_repo_id: the 96-char limit applies per
+    # segment (repo name / namespace), not to the whole "namespace/repo_name"
+    # string, so long-but-valid repo names are not falsely rejected.
     return all(
-        segment not in ("", ".", "..") and _VALID_REPO_ID_SEGMENT.fullmatch(segment) is not None
+        segment not in ("", ".", "..")
+        and len(segment) <= _MAX_REPO_ID_LENGTH
+        and _VALID_REPO_ID_SEGMENT.fullmatch(segment) is not None
         for segment in segments
     )
 
diff --git a/studio/backend/hub/utils/state_dir.py b/studio/backend/hub/utils/state_dir.py
index a304477a3d..183e934724 100644
--- a/studio/backend/hub/utils/state_dir.py
+++ b/studio/backend/hub/utils/state_dir.py
@@ -11,8 +11,9 @@ cache lifecycle. Two subdirectories:
         manifests/   <key>.json   per-download expected-files manifest
         cancelled/   <key>.json   per-download cancel marker
 
-The ``<key>`` mirrors HF's cache dir naming so a state file can be
-eyeballed next to the on-disk repo it describes:
+The ``<key>`` mirrors HF's cache dir naming while the resulting manifest,
+cancel-marker, and atomic-write temp filenames fit common filesystem basename
+limits. Very long repo IDs use a stable hash in the state key:
 
     models--<owner>--<name>                       full snapshot
     models--<owner>--<name>--variant--<variant>   GGUF variant
@@ -49,6 +50,11 @@ _MANIFESTS_SUBDIR = "manifests"
 _CANCELLED_SUBDIR = "cancelled"
 _WORKERS_SUBDIR = "workers"
 _SAFE_VARIANT_FRAGMENT = re.compile(r"^[a-z0-9._-]{1,64}$")
+_MAX_STATE_BASENAME_BYTES = 255
+_STATE_EXTENSION = ".json"
+# _atomic_write_json writes ".<target>.tmp-<8hex>" beside the final file.
+_ATOMIC_WRITE_TMP_OVERHEAD = len(".") + len(".tmp-") + 8
+_MAX_VARIANT_FRAGMENT_LENGTH = 64
 
 
 def state_root() -> Optional[Path]:
@@ -84,16 +90,35 @@ def repo_cache_basename(repo_type: RepoType, repo_id: str) -> str:
     return f"{repo_type}s--{repo_id.replace('/', '--')}".lower()
 
 
+def _filename_bytes(name: str) -> int:
+    return len(name.encode("utf-8"))
+
+
+def _state_filename_fits(entry_key: str) -> bool:
+    filename = f"{entry_key}{_STATE_EXTENSION}"
+    return _filename_bytes(filename) + _ATOMIC_WRITE_TMP_OVERHEAD <= _MAX_STATE_BASENAME_BYTES
+
+
+def _state_repo_key(repo_type: RepoType, repo_id: str) -> str:
+    base = repo_cache_basename(repo_type, repo_id)
+    variant_prefix = f"{base}--variant--"
+    longest_variant_key = f"{variant_prefix}{'x' * _MAX_VARIANT_FRAGMENT_LENGTH}"
+    if _state_filename_fits(longest_variant_key):
+        return base
+    digest = hashlib.sha256(base.encode("utf-8")).hexdigest()[:32]
+    return f"{repo_type}s--sha256-{digest}"
+
+
 def variant_filename_prefix(repo_type: RepoType, repo_id: str) -> str:
     """Lowercased prefix every variant-keyed state file for this repo shares.
 
     The single source the download_manifest enumerators match against, so the
     scheme in :func:`_entry_key` cannot drift from them silently."""
-    return f"{repo_cache_basename(repo_type, repo_id)}--variant--"
+    return f"{_state_repo_key(repo_type, repo_id)}--variant--"
 
 
 def _entry_key(repo_type: RepoType, repo_id: str, variant: Optional[str]) -> str:
-    base = repo_cache_basename(repo_type, repo_id)
+    base = _state_repo_key(repo_type, repo_id)
     if not variant:
         return base
     normalized_variant = variant.strip().lower()
diff --git a/studio/backend/models/inference.py b/studio/backend/models/inference.py
index 0f27b695fe..53b0f14b09 100644
--- a/studio/backend/models/inference.py
+++ b/studio/backend/models/inference.py
@@ -1533,12 +1533,41 @@ class AnthropicToolResultBlock(BaseModel):
     tool_use_id: str
     content: Union[str, list] = ""
 
+    @field_validator("content", mode = "before")
+    @classmethod
+    def _coerce_null_content(cls, v):
+        # Some clients send null content for an empty tool result; the str|list
+        # union would 400 on it, so treat null as "".
+        return "" if v is None else v
+
+
+# Block types the converter translates explicitly. Anything else (thinking /
+# redacted_thinking, a provider block a resumed session replays, or a future type)
+# is accepted as an unknown block and dropped by the converter, rather than 400-ing
+# the whole request on strict validation.
+_KNOWN_ANTHROPIC_BLOCK_TYPES = frozenset({"text", "image", "tool_use", "tool_result"})
+
+
+class AnthropicUnknownBlock(BaseModel):
+    type: str
+    model_config = {"extra": "allow"}
+
+    @field_validator("type")
+    @classmethod
+    def _only_unknown_types(cls, v):
+        # Known types parse as their typed models above (so a malformed known block
+        # still fails cleanly); this fallback only catches the rest.
+        if v in _KNOWN_ANTHROPIC_BLOCK_TYPES:
+            raise ValueError("known block type handled by its typed model")
+        return v
+
 
 AnthropicContentBlock = Union[
     AnthropicTextBlock,
     AnthropicImageBlock,
     AnthropicToolUseBlock,
     AnthropicToolResultBlock,
+    AnthropicUnknownBlock,
 ]
 
 
@@ -1583,6 +1612,40 @@ class AnthropicMessage(BaseModel):
     role: Literal["user", "assistant"]
     content: Union[str, list[AnthropicContentBlock]]
 
+    @model_validator(mode = "before")
+    @classmethod
+    def _normalize_content(cls, data):
+        # Role-aware leniency that never silently drops real user input:
+        #  - assistant: a resumed tool-only turn's null content -> "" (str|list would
+        #    400 on null; "" keeps the converter's `for block in content` safe).
+        #    Unknown blocks (thinking / future types) validate via
+        #    AnthropicUnknownBlock and are dropped by the converter.
+        #  - user: keep strict. Null user content stays None so str|list rejects it
+        #    (400) rather than forwarding an empty prompt; and reject block types the
+        #    converter cannot translate, since it silently skips unknown user blocks
+        #    -- a user turn made only of them would validate yet send no content
+        #    (silent data loss).
+        if not isinstance(data, dict):
+            return data
+        content = data.get("content")
+        if data.get("role") == "assistant":
+            # Coerce only an explicit null (resumed tool-only turn). A missing
+            # content key stays malformed so the required-field check still 400s.
+            if "content" in data and content is None:
+                return {**data, "content": ""}
+            return data
+        if isinstance(content, list):
+            for block in content:
+                btype = (
+                    block.get("type") if isinstance(block, dict) else getattr(block, "type", None)
+                )
+                # Guard the value: a non-string type is unsupported too, and a
+                # membership test on an unhashable value would raise TypeError
+                # (escaping as a 500 instead of a clean 400).
+                if not isinstance(btype, str) or btype not in _KNOWN_ANTHROPIC_BLOCK_TYPES:
+                    raise ValueError(f"unsupported content block type {btype!r} in a user message")
+        return data
+
 
 class AnthropicTool(BaseModel):
     # Client tools have input_schema; server tools may only have type/name.
diff --git a/studio/backend/routes/data_recipe/seed.py b/studio/backend/routes/data_recipe/seed.py
index 57a291291e..a5b75b7335 100644
--- a/studio/backend/routes/data_recipe/seed.py
+++ b/studio/backend/routes/data_recipe/seed.py
@@ -10,6 +10,7 @@ import binascii
 import json
 import os
 import re
+import shutil
 from itertools import islice
 from pathlib import Path
 from typing import Any
@@ -59,6 +60,9 @@ UNSTRUCTURED_ALLOWED_EXTS = {".pdf", ".docx", ".txt", ".md"}
 SEED_UPLOAD_DIR = seed_uploads_root()
 UNSTRUCTURED_UPLOAD_ROOT = unstructured_uploads_root()
 _SAFE_ID_RE = re.compile(r"^[a-zA-Z0-9_-]+$")
+# Frontend-generated upload namespace (UUID4 hex). Legacy node ids (n1, ...)
+# never match: those directories can be shared by several recipes.
+_UPLOAD_UID_RE = re.compile(r"^[0-9a-f]{32}$")
 
 
 def _validate_safe_id(value: str, label: str) -> str:
@@ -580,6 +584,39 @@ async def remove_unstructured_file(block_id: str, file_id: str):
     return {"status": "ok"}
 
 
+@router.delete("/seed/unstructured-block/{block_id}")
+async def remove_unstructured_block(block_id: str):
+    """Delete a block's upload directory; files on disk still count toward its quota.
+
+    Only uid-namespaced directories may be bulk-deleted: they have exactly one
+    owning block. Legacy node-id directories (n1, ...) can be shared by other
+    recipes, so they are managed file-by-file instead.
+    """
+    _validate_safe_id(block_id, "block_id")
+    if not _UPLOAD_UID_RE.match(block_id):
+        raise HTTPException(400, "Invalid block_id: only uid-namespaced blocks can be deleted")
+
+    block_dir = (UNSTRUCTURED_UPLOAD_ROOT / block_id).resolve()
+    if not block_dir.is_relative_to(UNSTRUCTURED_UPLOAD_ROOT.resolve()):
+        raise HTTPException(400, "Invalid block_id: outside upload root")
+    if not block_dir.exists():
+        return {"status": "ok", "deleted": False}
+
+    try:
+        shutil.rmtree(block_dir)
+    except OSError as exc:
+        raise log_and_http_error(
+            exc,
+            500,
+            "failed to delete uploaded files",
+            event = "data_recipe.seed.unstructured_block_delete_failed",
+            log = logger,
+        ) from exc
+    if block_dir.exists():
+        raise HTTPException(500, "failed to delete uploaded files")
+    return {"status": "ok", "deleted": True}
+
+
 @router.post("/seed/inspect-upload", response_model = SeedInspectResponse)
 def inspect_seed_upload(payload: SeedInspectUploadRequest) -> SeedInspectResponse:
     if payload.file_ids is not None:
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 5332037e0d..ec5d309810 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -13,7 +13,7 @@ from pathlib import Path
 from fastapi import APIRouter, Depends, HTTPException, Request, status
 from fastapi.responses import StreamingResponse, JSONResponse, Response
 from starlette.requests import ClientDisconnect
-from typing import Any, List, Optional, Union
+from typing import Any, Callable, List, Optional, Union
 import json
 import httpx
 from loggers import get_logger
@@ -28,6 +28,16 @@ import re as _re
 from utils.models import extract_model_size_b as _extract_model_size_b
 
 from utils.api_errors import openai_error_body, anthropic_error_body
+from core.inference.llama_admission import (
+    LlamaAdmissionCancelled,
+    LlamaAdmissionConfig,
+    LlamaAdmissionLease,
+    LlamaAdmissionQueueFull,
+    LlamaAdmissionReservation,
+    LlamaAdmissionTimeout,
+    get_llama_admission_queue,
+    llama_admission_config_from_env,
+)
 
 
 def _positive_int_or_none(value: Any) -> Optional[int]:
@@ -40,6 +50,43 @@ def _positive_int_or_none(value: Any) -> Optional[int]:
     return value_int if value_int > 0 else None
 
 
+def _nonnegative_int_or_none(value: Any) -> Optional[int]:
+    if isinstance(value, bool):
+        return None
+    try:
+        value_int = int(value)
+    except (TypeError, ValueError):
+        return None
+    return value_int if value_int >= 0 else None
+
+
+_MLX_MPI_DISTRIBUTED_ENV_PAIRS = (
+    ("OMPI_COMM_WORLD_RANK", "OMPI_COMM_WORLD_SIZE"),
+    ("PMI_RANK", "PMI_SIZE"),
+    ("PMIX_RANK", "PMIX_SIZE"),
+    ("MPI_RANK", "MPI_WORLD_SIZE"),
+    ("MV2_COMM_WORLD_RANK", "MV2_COMM_WORLD_SIZE"),
+)
+
+
+def _mlx_distributed_launch_detected() -> bool:
+    if _nonnegative_int_or_none(os.environ.get("MLX_RANK")) is not None:
+        world_size = _positive_int_or_none(os.environ.get("MLX_WORLD_SIZE"))
+        if world_size is not None and world_size > 1:
+            return True
+        return bool(
+            os.environ.get("MLX_HOSTFILE")
+            or os.environ.get("MLX_IBV_DEVICES")
+            or os.environ.get("MLX_JACCL_COORDINATOR")
+            or (os.environ.get("NCCL_HOST_IP") and os.environ.get("NCCL_PORT"))
+        )
+    return any(
+        _nonnegative_int_or_none(os.environ.get(rank_env)) is not None
+        and (_positive_int_or_none(os.environ.get(size_env)) or 0) > 1
+        for rank_env, size_env in _MLX_MPI_DISTRIBUTED_ENV_PAIRS
+    )
+
+
 def _install_httpcore_asyncgen_silencer() -> None:
     """Silence benign httpx/httpcore asyncgen GC noise on Python 3.13.
 
@@ -231,10 +278,87 @@ def _effective_max_tokens(payload):
     )
 
 
+_OPENAI_COMPAT_STREAM_STALL_TIMEOUT_ENV = "UNSLOTH_OPENAI_COMPAT_STREAM_STALL_TIMEOUT"
+
+
+def _positive_float_env(env_name: str, default):
+    """Parse a positive float from an env var. A parseable non-positive value
+    returns ``None`` (0 disables the guarded feature); only unparseable or unset
+    values fall back to ``default``."""
+    raw_value = os.environ.get(env_name)
+    if raw_value is None or not raw_value.strip():
+        return default
+    try:
+        value = float(raw_value.strip())
+    except ValueError:
+        return default
+    return value if value > 0 else None
+
+
+def _effective_openai_max_tokens_from_values(max_tokens, max_completion_tokens = None):
+    """Resolve the OpenAI-compatible generation cap from raw request values.
+
+    Prefers ``max_completion_tokens`` over the deprecated ``max_tokens``, and
+    returns ``None`` when both are omitted so callers keep their context-window
+    default (OpenAI treats an omitted cap as bounded only by the context
+    window). Explicit client caps pass through unchanged.
+    """
+
+    def _validate_explicit(value, param: str):
+        if value is None:
+            return None
+        if isinstance(value, bool) or not isinstance(value, int):
+            raise HTTPException(
+                status_code = 400,
+                detail = openai_error_body(
+                    f"'{param}' must be an integer.",
+                    status = 400,
+                    code = "invalid_type",
+                    param = param,
+                ),
+            )
+        # The legacy completions spec declares ``minimum: 0`` for max_tokens,
+        # so 0 is a valid (if degenerate) cap and only negatives are rejected.
+        # The chat fields never reach here with 0 (pydantic enforces ge=1).
+        if value < 0:
+            raise HTTPException(
+                status_code = 400,
+                detail = openai_error_body(
+                    f"'{param}' must be at least 0.",
+                    status = 400,
+                    code = "invalid_value",
+                    param = param,
+                ),
+            )
+        return value
+
+    max_tokens = _validate_explicit(max_tokens, "max_tokens")
+    max_completion_tokens = _validate_explicit(max_completion_tokens, "max_completion_tokens")
+    return max_completion_tokens if max_completion_tokens is not None else max_tokens
+
+
+def _effective_openai_max_tokens(payload):
+    return _effective_openai_max_tokens_from_values(
+        getattr(payload, "max_tokens", None),
+        getattr(payload, "max_completion_tokens", None),
+    )
+
+
 def _wants_multiple_choices(payload) -> bool:
     return (payload.n or 1) > 1
 
 
+def _has_openai_tool_history(messages) -> bool:
+    for message in messages or []:
+        if isinstance(message, dict):
+            if message.get("role") == "tool" or message.get("tool_calls"):
+                return True
+            continue
+        if getattr(message, "role", None) == "tool" or getattr(message, "tool_calls", None):
+            return True
+    return False
+
+
 def _raise_unsupported_openai_parameter(param: str, message: str) -> None:
     raise HTTPException(
         status_code = 400,
@@ -294,6 +418,14 @@ def _openai_stream_error_chunk(exc) -> dict:
     return openai_error_body(_friendly_error(exc), status = 500)
 
 
+def _openai_stream_error_sse(error: dict) -> str:
+    return f"data: {json.dumps(error)}\n\ndata: [DONE]\n\n"
+
+
+def _openai_stream_error_sse_bytes(error: dict) -> bytes:
+    return _openai_stream_error_sse(error).encode("utf-8")
+
+
 def _openai_passthrough_error(status_code, text) -> "HTTPException":
     """HTTPException for a non-200 upstream response on the OpenAI passthrough
     (tools / response_format). An over-context upstream error is mapped to a 400
@@ -508,20 +640,62 @@ def _drop_parallel_tool_call_deltas(chunk) -> bool:
     return changed
 
 
-def _cap_parallel_tool_calls_sse_line(raw_line: str) -> str:
-    """Drop tool_call deltas whose index >= 1 from one streamed OpenAI SSE
-    ``data:`` line so only the first tool call survives (parallel_tool_calls=false,
-    best-effort). Non-tool / unparseable payloads are returned byte-for-byte."""
-    payload = raw_line[len("data: ") :]
+def _add_empty_content_to_reasoning_deltas(chunk: dict) -> bool:
+    """Make reasoning-only deltas palatable to strict OpenAI adapters.
+
+    Some clients built on OpenAI-compatible streams ignore or reject chunks whose
+    delta only contains non-standard ``reasoning_content``. Preserve that field,
+    but add an empty standard ``content`` member so the chunk is still a valid
+    text-delta shape and downstream parsers keep the stream alive.
+    """
+    changed = False
+    choices = chunk.get("choices")
+    if not isinstance(choices, list):
+        return False
+    for choice in choices:
+        if not isinstance(choice, dict):
+            continue
+        delta = choice.get("delta")
+        if not isinstance(delta, dict):
+            continue
+        if "reasoning_content" in delta and "content" not in delta:
+            delta["content"] = ""
+            changed = True
+    return changed
+
+
+def _normalize_openai_passthrough_sse_line(
+    raw_line: str, *, cap_parallel_tool_calls: bool = False
+) -> str:
+    """Normalize one passthrough OpenAI SSE ``data:`` line before relaying.
+
+    The function is intentionally narrow: it leaves comments, blank events,
+    ``[DONE]``, and unparseable upstream bytes untouched; parsed chunks are
+    re-serialized only when a compatibility mutation is actually required.
+    """
+    if not raw_line.startswith("data:"):
+        return raw_line
+    # Both mutations key off JSON object keys, so a line without either quoted
+    # key can never change; skip the parse on the per-token common case.
+    if '"reasoning_content"' not in raw_line and not (
+        cap_parallel_tool_calls and '"tool_calls"' in raw_line
+    ):
+        return raw_line
+    payload = raw_line[len("data:") :].lstrip()
     if payload.strip() in ("", "[DONE]"):
         return raw_line
     try:
         obj = json.loads(payload)
     except Exception:
         return raw_line
-    if not _drop_parallel_tool_call_deltas(obj):
+    if not isinstance(obj, dict):
         return raw_line
-    return "data: " + json.dumps(obj, separators = (",", ":"))
+    changed = _add_empty_content_to_reasoning_deltas(obj)
+    if cap_parallel_tool_calls and _drop_parallel_tool_call_deltas(obj):
+        changed = True
+    if not changed:
+        return raw_line
+    return "data: " + json.dumps(obj, separators = (",", ":"), ensure_ascii = False)
 
 
 def _prompt_tokens_details(upstream):
@@ -538,6 +712,49 @@ def _wants_stream_usage(payload) -> bool:
     return bool((payload.stream_options or {}).get("include_usage"))
 
 
+_OPENAI_PASSTHROUGH_TERMINAL_GRACE_S = 2.0
+_SSE_DONE_LINE = "data: [DONE]"
+
+
+def _openai_passthrough_sse_line_terminal_state(raw_line: str) -> Optional[str]:
+    """Classify OpenAI-compatible chat stream terminal markers.
+
+    Some llama-server builds can emit the logical final chunk (``finish_reason``)
+    and optional usage chunk, then keep the HTTP stream open without sending the
+    OpenAI ``data: [DONE]`` sentinel. Classifying those chunks lets Studio close
+    the client stream promptly while preserving an optional trailing usage chunk.
+    """
+    if not raw_line.startswith("data:"):
+        return None
+    data_str = raw_line[5:].lstrip()
+    if data_str == "[DONE]":
+        return "done"
+    try:
+        data = json.loads(data_str)
+    except json.JSONDecodeError:
+        return None
+    return _openai_passthrough_terminal_state_from_data(data)
+
+
+def _openai_passthrough_terminal_state_from_data(data) -> Optional[str]:
+    """Dict-level core of ``_openai_passthrough_sse_line_terminal_state`` for
+    callers that already parsed the chunk (avoids a re-parse per relayed line)."""
+    if not isinstance(data, dict):
+        return None
+    if _monitor_openai_error_message(data):
+        return "error"
+    choices = data.get("choices")
+    if isinstance(choices, list):
+        if not choices and isinstance(data.get("usage"), dict):
+            return "usage"
+        for choice in choices:
+            if isinstance(choice, dict) and choice.get("finish_reason") is not None:
+                return "finish"
+    elif isinstance(data.get("usage"), dict):
+        return "usage"
+    return None
+
+
 def _openai_stream_usage_chunk(
     payload, completion_id, created, model_name, stream_usage, stream_timings
 ):
@@ -604,12 +821,16 @@ def _chat_content_chunk(completion_id, created, model_name, text) -> str:
 
 
 def _chat_reasoning_chunk(completion_id, created, model_name, text) -> str:
-    """Like ``_chat_content_chunk`` but on ``reasoning_content`` (renders the UI thinking block)."""
+    """Like ``_chat_content_chunk`` but on ``reasoning_content`` (renders the UI thinking block).
+
+    Carries ``content: ""`` alongside, like the GGUF and passthrough paths, so
+    strict OpenAI adapters don't drop the reasoning-only delta.
+    """
     return _chat_chunk_sse(
         completion_id,
         created,
         model_name,
-        delta = ChoiceDelta(reasoning_content = text),
+        delta = ChoiceDelta(content = "", reasoning_content = text),
         finish_reason = None,
     )
 
@@ -844,8 +1065,11 @@ def _llama_streaming_generation_timeout() -> httpx.Timeout:
 
 
 def _set_stream_response_read_timeout(
-    response: httpx.Response, read_timeout_s: float = _DEFAULT_STREAM_STALL_TIMEOUT_S
+    response: httpx.Response, read_timeout_s: Optional[float] = _DEFAULT_STREAM_STALL_TIMEOUT_S
 ) -> None:
+    # ``read_timeout_s = None`` clears httpx's read timeout (wait indefinitely),
+    # used when the stall guard is disabled so a stale first-token deadline
+    # can't keep timing out post-first-chunk gaps.
     try:
         timeout_ext = response.request.extensions.get("timeout")
         if isinstance(timeout_ext, dict):
@@ -856,6 +1080,295 @@ def _set_stream_response_read_timeout(
 
 _STREAM_DISCONNECT_POLL_TIMEOUT_S = 0.25
 _OPENAI_PASSTHROUGH_PREHEADER_STATUS_WINDOW_S = 0.1
+_OPENAI_PASSTHROUGH_PENDING_RESPONSE_KEEPALIVE_S = 5.0
+_OPENAI_PASSTHROUGH_SSE_KEEPALIVE = ": keep-alive\n\n"
+_OPENAI_LLAMA_ADMISSION_POLL_S = 0.25
+
+
+def _openai_llama_admission_capacity(request: Optional[Request], llama_backend = None) -> int:
+    """Serving slots available for one local llama-server backend.
+
+    The loaded backend is the source of truth because it may have reduced
+    ``--parallel`` at load time to keep the model on GPU. The app state is a
+    launch-intent fallback for tests and for the short window before a backend
+    reports its committed runtime slots.
+    """
+    slots = _positive_int_or_none(getattr(llama_backend, "effective_parallel_slots", None))
+    if slots is not None:
+        return slots
+    try:
+        slots = getattr(request.app.state, "llama_parallel_slots", None)
+    except Exception:
+        slots = None
+    return _positive_int_or_none(slots) or 1
+
+
+def _openai_llama_admission_reserve(
+    *, request: Optional[Request], llama_backend
+) -> tuple[LlamaAdmissionReservation, LlamaAdmissionConfig]:
+    config = llama_admission_config_from_env()
+    capacity = _openai_llama_admission_capacity(request, llama_backend)
+    key = str(getattr(llama_backend, "base_url", "llama-server"))
+    reservation = get_llama_admission_queue(key).reserve(
+        capacity = capacity,
+        config = config,
+    )
+    return reservation, config
+
+
+def _openai_admission_request_path(request: Optional[Request]) -> Optional[str]:
+    try:
+        return str(request.url.path) if request is not None else None
+    except Exception:
+        return None
+
+
+def _openai_admission_log(
+    event: str,
+    reservation: Optional[LlamaAdmissionReservation] = None,
+    *,
+    snapshot = None,
+    request: Optional[Request],
+    mode: str,
+    wait_started_at: Optional[float] = None,
+    completion_id: Optional[str] = None,
+    level: str = "debug",
+) -> None:
+    if snapshot is None and reservation is not None:
+        snapshot = reservation.snapshot_now()
+    wait_ms = None
+    if wait_started_at is not None:
+        wait_ms = int(max(0.0, time.monotonic() - wait_started_at) * 1000)
+    log = getattr(logger, level, logger.debug)
+    log(
+        "openai admission %s: mode=%s path=%s completion_id=%s capacity=%s active=%s queued=%s wait_ms=%s",
+        event,
+        mode,
+        _openai_admission_request_path(request),
+        completion_id,
+        getattr(snapshot, "capacity", None),
+        getattr(snapshot, "active", None),
+        getattr(snapshot, "queued", None),
+        wait_ms,
+    )
+
+
+def _openai_admission_error_body(exc: Exception, *, status_code: int) -> dict:
+    snapshot = getattr(exc, "snapshot", None)
+    message = str(exc)
+    if snapshot is not None:
+        message = (
+            f"{message} "
+            f"(active={snapshot.active}, queued={snapshot.queued}, capacity={snapshot.capacity})"
+        )
+    return openai_error_body(message, status = status_code)
+
+
+def _openai_admission_http_exception(exc: Exception, *, status_code: int) -> HTTPException:
+    return HTTPException(
+        status_code = status_code,
+        detail = _openai_admission_error_body(exc, status_code = status_code),
+    )
+
+
+def _openai_admission_timeout_error(
+    reservation: LlamaAdmissionReservation,
+) -> LlamaAdmissionTimeout:
+    return LlamaAdmissionTimeout(
+        "Timed out waiting for an available local llama-server generation slot",
+        snapshot = reservation.snapshot_now(),
+    )
+
+
+def _openai_admission_cancelled_error(
+    reservation: LlamaAdmissionReservation,
+) -> LlamaAdmissionCancelled:
+    return LlamaAdmissionCancelled(
+        "Client disconnected before an upstream llama-server generation slot was available",
+        snapshot = reservation.snapshot_now(),
+    )
+
+
+async def _raise_if_openai_admission_cancelled(
+    reservation: LlamaAdmissionReservation, *, request: Optional[Request], cancel_event
+) -> None:
+    if reservation.is_cancelled:
+        raise _openai_admission_cancelled_error(reservation)
+    if await _preheader_cancelled(cancel_event, request):
+        reservation.cancel()
+        raise _openai_admission_cancelled_error(reservation)
+
+
+async def _wait_for_openai_admission_non_streaming(
+    reservation: LlamaAdmissionReservation,
+    config: LlamaAdmissionConfig,
+    *,
+    request: Optional[Request],
+    cancel_event,
+) -> LlamaAdmissionLease:
+    lease = reservation.lease_nowait()
+    if lease is not None:
+        try:
+            await _raise_if_openai_admission_cancelled(
+                reservation,
+                request = request,
+                cancel_event = cancel_event,
+            )
+        except asyncio.CancelledError:
+            lease.release()
+            raise
+        except LlamaAdmissionCancelled:
+            lease.release()
+            raise
+        return lease
+    await _raise_if_openai_admission_cancelled(
+        reservation,
+        request = request,
+        cancel_event = cancel_event,
+    )
+    deadline = None if config.queue_timeout_s is None else time.monotonic() + config.queue_timeout_s
+    try:
+        while True:
+            await _raise_if_openai_admission_cancelled(
+                reservation,
+                request = request,
+                cancel_event = cancel_event,
+            )
+            lease = reservation.lease_nowait()
+            if lease is not None:
+                try:
+                    await _raise_if_openai_admission_cancelled(
+                        reservation,
+                        request = request,
+                        cancel_event = cancel_event,
+                    )
+                except asyncio.CancelledError:
+                    lease.release()
+                    raise
+                except LlamaAdmissionCancelled:
+                    lease.release()
+                    raise
+                return lease
+            wait_s = _OPENAI_LLAMA_ADMISSION_POLL_S
+            if deadline is not None:
+                remaining_s = deadline - time.monotonic()
+                if remaining_s <= 0:
+                    reservation.cancel()
+                    raise _openai_admission_timeout_error(reservation)
+                wait_s = min(wait_s, max(remaining_s, 0.001))
+            try:
+                lease = await reservation.wait(wait_s)
+            except asyncio.TimeoutError:
+                continue
+            if lease is not None:
+                return lease
+            await _raise_if_openai_admission_cancelled(
+                reservation,
+                request = request,
+                cancel_event = cancel_event,
+            )
+    except asyncio.CancelledError:
+        reservation.cancel()
+        raise
+
+
+async def _openai_admission_wait_stream_chunks(
+    reservation: LlamaAdmissionReservation,
+    config: LlamaAdmissionConfig,
+    *,
+    request: Optional[Request],
+    cancel_event,
+):
+    lease = reservation.lease_nowait()
+    if lease is not None:
+        yield lease
+        return
+
+    await _raise_if_openai_admission_cancelled(
+        reservation,
+        request = request,
+        cancel_event = cancel_event,
+    )
+    deadline = None if config.queue_timeout_s is None else time.monotonic() + config.queue_timeout_s
+    keepalive_interval_s = max(0.001, config.keepalive_interval_s)
+    next_keepalive_at = time.monotonic() + keepalive_interval_s
+    try:
+        while True:
+            await _raise_if_openai_admission_cancelled(
+                reservation,
+                request = request,
+                cancel_event = cancel_event,
+            )
+            lease = reservation.lease_nowait()
+            if lease is not None:
+                yield lease
+                return
+
+            now = time.monotonic()
+            wait_s = min(_OPENAI_LLAMA_ADMISSION_POLL_S, max(next_keepalive_at - now, 0.001))
+            if deadline is not None:
+                remaining_s = deadline - now
+                if remaining_s <= 0:
+                    reservation.cancel()
+                    raise _openai_admission_timeout_error(reservation)
+                wait_s = min(wait_s, max(remaining_s, 0.001))
+            try:
+                lease = await reservation.wait(wait_s)
+            except asyncio.TimeoutError:
+                lease = None
+            if lease is not None:
+                yield lease
+                return
+            await _raise_if_openai_admission_cancelled(
+                reservation,
+                request = request,
+                cancel_event = cancel_event,
+            )
+            now = time.monotonic()
+            if now >= next_keepalive_at:
+                next_keepalive_at = now + keepalive_interval_s
+                yield _OPENAI_PASSTHROUGH_SSE_KEEPALIVE
+    except asyncio.CancelledError:
+        reservation.cancel()
+        raise
+
+
+async def _close_openai_admitted_stream_iterator(iterator, *, cancelled: bool) -> None:
+    if iterator is None:
+        return
+    if cancelled:
+        athrow = getattr(iterator, "athrow", None)
+        if athrow is not None:
+            try:
+                await athrow(asyncio.CancelledError())
+            except (asyncio.CancelledError, StopAsyncIteration, RuntimeError):
+                return
+    aclose = getattr(iterator, "aclose", None)
+    if aclose is not None:
+        await aclose()
+
+
+def _openai_compat_stream_stall_timeout():
+    """Max silent gap after an OpenAI passthrough stream has produced data.
+
+    If the socket goes silent after valid SSE data, this bounds how long the
+    client is kept open. Defaults to the backend-wide stall timeout so this
+    path stalls out like every sibling stream; set the env var to tighten it
+    for local serving, or to 0 to disable the guard.
+    """
+    return _positive_float_env(
+        _OPENAI_COMPAT_STREAM_STALL_TIMEOUT_ENV,
+        _DEFAULT_STREAM_STALL_TIMEOUT_S,
+    )
+
+
+def _openai_passthrough_upstream_headers(*, llama_backend = None) -> dict:
+    headers = {}
+    auth_headers = getattr(llama_backend, "_auth_headers", None)
+    if isinstance(auth_headers, dict):
+        headers.update(auth_headers)
+    headers["Connection"] = "close"
+    return headers
 
 
 class _CompatSameTaskTimeout:
@@ -983,7 +1496,8 @@ async def _aclose_stream_resources(
     """Tear down an httpx streaming generator's resources in the required order:
     cancel + await each watcher task, then aclose() the byte/line iterator, the
     response, and the client. Each step swallows its own exceptions so teardown
-    always completes. See _anthropic_passthrough_stream for the ordering rationale."""
+    always completes; a close-time CancelledError is re-raised only after every
+    step has run. See _anthropic_passthrough_stream for the ordering rationale."""
     for watcher in watchers:
         if watcher is not None:
             watcher.cancel()
@@ -991,21 +1505,30 @@ async def _aclose_stream_resources(
                 await watcher
             except (asyncio.CancelledError, Exception):
                 pass
+    close_cancelled = False
     if iterator is not None:
         try:
             await iterator.aclose()
+        except asyncio.CancelledError:
+            close_cancelled = True
         except Exception:
             pass
     if resp is not None:
         try:
             await resp.aclose()
+        except asyncio.CancelledError:
+            close_cancelled = True
         except Exception:
             pass
     if client is not None:
         try:
             await client.aclose()
+        except asyncio.CancelledError:
+            close_cancelled = True
         except Exception:
             pass
+    if close_cancelled:
+        raise asyncio.CancelledError()
 
 
 async def _preheader_cancelled(cancel_event = None, request: Optional[Request] = None) -> bool:
@@ -1028,6 +1551,7 @@ async def _send_stream_with_preheader_cancel(
     req: httpx.Request,
     cancel_event = None,
     request: Optional[Request] = None,
+    mark_cancel_on_cancel: bool = True,
 ) -> Optional[httpx.Response]:
     if cancel_event is None and request is None:
         return await client.send(req, stream = True)
@@ -1059,7 +1583,7 @@ async def _send_stream_with_preheader_cancel(
         await _stop_send_task()
         return None
     except asyncio.CancelledError:
-        if cancel_event is not None:
+        if mark_cancel_on_cancel and cancel_event is not None:
             cancel_event.set()
         await _stop_send_task()
         raise
@@ -1078,11 +1602,19 @@ async def _aiter_llama_stream_items(
     request: Optional[Request] = None,
     first_token_deadline: Optional[float] = None,
     response: Optional[httpx.Response] = None,
-    post_first_item_read_timeout_s: Optional[float] = _DEFAULT_STREAM_STALL_TIMEOUT_S,
+    post_first_item_read_timeout_s: Optional[
+        Union[float, Callable[[], Optional[float]]]
+    ] = _DEFAULT_STREAM_STALL_TIMEOUT_S,
 ):
     if first_token_deadline is None:
         first_token_deadline = time.monotonic() + _DEFAULT_FIRST_TOKEN_TIMEOUT_S
     last_item_at: Optional[float] = None
+
+    def _post_first_timeout_s() -> Optional[float]:
+        if callable(post_first_item_read_timeout_s):
+            return post_first_item_read_timeout_s()
+        return post_first_item_read_timeout_s
+
     while True:
         if cancel_event is not None and cancel_event.is_set():
             return
@@ -1103,15 +1635,14 @@ async def _aiter_llama_stream_items(
                 async with _same_task_timeout(remaining_s):
                     item = await async_iter.__anext__()
             else:
+                timeout_s = _post_first_timeout_s()
                 if (
                     request is not None
                     and response is not None
-                    and post_first_item_read_timeout_s is not None
+                    and timeout_s is not None
                     and last_item_at is not None
                 ):
-                    stall_remaining_s = post_first_item_read_timeout_s - (
-                        time.monotonic() - last_item_at
-                    )
+                    stall_remaining_s = timeout_s - (time.monotonic() - last_item_at)
                     if stall_remaining_s <= 0:
                         raise httpx.ReadTimeout("The model stopped producing tokens mid-response.")
                     _set_stream_response_read_timeout(response, stall_remaining_s)
@@ -1128,19 +1659,16 @@ async def _aiter_llama_stream_items(
                 if now >= first_token_deadline:
                     raise
                 continue
-            if (
-                request is not None
-                and post_first_item_read_timeout_s is not None
-                and now - last_item_at < post_first_item_read_timeout_s
-            ):
+            timeout_s = _post_first_timeout_s()
+            if request is not None and timeout_s is not None and now - last_item_at < timeout_s:
                 continue
             raise httpx.ReadTimeout("The model stopped producing tokens mid-response.")
-        if (
-            last_item_at is None
-            and response is not None
-            and post_first_item_read_timeout_s is not None
-        ):
-            _set_stream_response_read_timeout(response, post_first_item_read_timeout_s)
+        if last_item_at is None and response is not None:
+            # The first-token read deadline no longer applies once a chunk has
+            # arrived: switch to the stall timeout, or clear the read timeout
+            # entirely when the stall guard is disabled (callable returns None)
+            # so a long gap can't trip the stale first-token deadline.
+            _set_stream_response_read_timeout(response, _post_first_timeout_s())
         last_item_at = time.monotonic()
         yield item
 
@@ -1519,6 +2047,20 @@ def _effective_enable_tools(payload) -> Optional[bool]:
     return policy if policy is not None else payload.enable_tools
 
 
+def _explicit_studio_tool_loop_requested(payload) -> bool:
+    """True when the request itself asks Studio to execute local tools.
+
+    Process-wide CLI policy can default Studio's tool loop on for ordinary chat,
+    but it must not steal OpenAI-compatible client tools or response_format
+    requests from the llama-server passthrough path. A policy of ``False``
+    (--disable-tools) vetoes even an explicit ``enable_tools: true`` ask.
+    """
+    from state.tool_policy import get_tool_policy
+
+    policy = get_tool_policy()
+    return policy is not False and (payload.enable_tools is True or bool(payload.mcp_enabled))
+
+
 # Cancel registry. Proxies (e.g. Colab) can swallow client fetch aborts so
 # is_disconnected() never fires. POST /inference/cancel looks up in-flight
 # cancel_events here by cancel_id (per-run) or session_id / completion_id
@@ -1659,6 +2201,39 @@ async def _await_disconnect_then_cancel(request, cancel_event) -> None:
         return
 
 
+def _cancelable_nonstreaming_client() -> httpx.AsyncClient:
+    return httpx.AsyncClient(
+        limits = httpx.Limits(max_connections = 1, max_keepalive_connections = 0),
+        trust_env = False,
+    )
+
+
+async def _await_cancel_or_disconnect_then_close_client(
+    *, cancel_event, request: Optional[Request], client: httpx.AsyncClient
+) -> None:
+    """Close a dedicated non-streaming upstream client on cancel/disconnect.
+
+    The shared ``nonstreaming_client()`` is pooled, so cancelable generation calls
+    use a per-request client. Closing it interrupts a blocked llama-server
+    request without affecting unrelated pooled non-streaming calls.
+    """
+    try:
+        while True:
+            if cancel_event is not None and cancel_event.is_set():
+                break
+            if request is not None and await request.is_disconnected():
+                if cancel_event is not None:
+                    cancel_event.set()
+                break
+            await asyncio.sleep(0.1)
+        try:
+            await client.aclose()
+        except Exception:
+            pass
+    except asyncio.CancelledError:
+        return
+
+
 async def _stop_local_disconnect_cancel_watcher(watcher) -> None:
     watcher.cancel()
     try:
@@ -2769,6 +3344,21 @@ def _automatic_model_load_may_run() -> bool:
     return get_openai_auto_switch_enabled() or get_auto_unload_idle_seconds() > 0
 
 
+def _no_model_loaded_detail(base: str) -> str:
+    """Append a pointer to the opt-in auto-switch toggle to a "no model loaded"
+    error, but only when it's off. Auto-switch (default off) cold-loads a
+    requested downloaded GGUF, so an off toggle is the usual reason a request
+    naming a listed model still 400/503s; surface the fix. With it on the name
+    simply didn't resolve to a local GGUF, so the hint would mislead and is omitted."""
+    from utils.openai_auto_switch_settings import get_openai_auto_switch_enabled
+
+    if get_openai_auto_switch_enabled():
+        return base
+    return base + (
+        " Or enable Model auto-switch (Settings > API) to load a requested model automatically."
+    )
+
+
 async def _maybe_auto_switch_model(
     requested_model: Optional[str],
     fastapi_request: Request,
@@ -3038,10 +3628,14 @@ def _remote_gguf_companion_bytes(
         info = model_info(repo, token = hf_token, files_metadata = True)
         total = 0
         for sibling in info.siblings or []:
-            base = Path(sibling.rfilename or "").name.lower()
+            name = sibling.rfilename or ""
+            base = Path(name).name.lower()
             if not base.endswith(".gguf"):
                 continue
-            if base.startswith("mtp-") or (include_mmproj and "mmproj" in base):
+            # Root-level mtp- only: -hf auto-fetches the repo-root drafter, not
+            # the MTP/ subdir copies (which now share the mtp- prefix too).
+            is_root_mtp = "/" not in name and base.startswith("mtp-")
+            if is_root_mtp or (include_mmproj and "mmproj" in base):
                 total += getattr(sibling, "size", 0) or 0
         return total
     except Exception as e:
@@ -3426,6 +4020,15 @@ async def _load_model_impl(request: LoadRequest, fastapi_request: Request, curre
                 status_code = 400,
                 detail = "gpu_ids is not supported for GGUF models yet.",
             )
+        if not config.is_gguf and _mlx_distributed_launch_detected():
+            raise HTTPException(
+                status_code = 400,
+                detail = (
+                    "Studio does not support distributed MLX inference under "
+                    "mlx.launch. Use `mlx.launch ... unsloth chat` or run Studio "
+                    "without the distributed launcher."
+                ),
+            )
 
         # Effective quantization (LoRA can flip 4-bit -> 16-bit); guard + load reuse it.
         effective_load_in_4bit = _effective_load_in_4bit(config, request.load_in_4bit)
@@ -5863,7 +6466,7 @@ async def openai_chat_completions(
         if not backend.active_model_name:
             raise HTTPException(
                 status_code = 400,
-                detail = "No model loaded. Call POST /inference/load first.",
+                detail = _no_model_loaded_detail("No model loaded. Call POST /inference/load first."),
             )
         # Clean public id so the response never echoes a local path; the audio
         # branch below receives this sanitized label too.
@@ -6040,11 +6643,11 @@ async def openai_chat_completions(
     # unaware of `role="tool"` messages and assistant messages that only
     # carry `tool_calls` (content=None) — both of which are valid in
     # multi-turn client-side tool loops.
-    effective_max_tokens = _effective_max_tokens(payload)
+    effective_max_tokens = _effective_openai_max_tokens(payload)
 
     normalized_stop = _normalize_stop_sequences(payload.stop)
 
-    _has_tool_messages = any(m.role == "tool" or m.tool_calls for m in payload.messages)
+    _has_tool_messages = _has_openai_tool_history(payload.messages)
     # Route guided-decoding requests through the verbatim passthrough so
     # ``response_format`` (JSON schema) reaches llama-server and the model's
     # GBNF-constrained output comes back unmodified. The non-passthrough GGUF
@@ -6053,12 +6656,41 @@ async def openai_chat_completions(
     # free-form sampling. Guided decoding does not require ``supports_tools`` --
     # the grammar machinery is independent of tool-call parsing.
     _has_response_format = _extract_response_format(payload) is not None
-    _tools_passthrough = llama_backend.supports_tools and (
-        (payload.tools and len(payload.tools) > 0) or _has_tool_messages
+    _has_tool_catalog = bool(payload.tools and len(payload.tools) > 0)
+    _has_active_tool_catalog = _has_tool_catalog and payload.tool_choice != "none"
+    _has_client_tool_contract = _has_active_tool_catalog or _has_tool_messages
+    # The Studio tool loop needs a tool-capable backend, so a request that asks
+    # for it on a backend that can't run it (DiffusionGemma forces supports_tools
+    # off) must not steal client tools from the passthrough (#6851).
+    _studio_tool_loop_requested = (
+        _explicit_studio_tool_loop_requested(payload) and llama_backend.supports_tools
     )
+    _client_disabled_tool_calls = payload.tool_choice == "none" and not _studio_tool_loop_requested
+    _supports_tool_passthrough = getattr(
+        llama_backend, "supports_tool_passthrough", llama_backend.supports_tools
+    )
+    _tools_passthrough = _supports_tool_passthrough and _has_client_tool_contract
     if (
         using_gguf
-        and not _effective_enable_tools(payload)
+        and not _studio_tool_loop_requested
+        and _has_client_tool_contract
+        and not _supports_tool_passthrough
+    ):
+        raise _reject(
+            400,
+            openai_error_body(
+                (
+                    "Client-supplied tools or tool-call history require a GGUF chat template "
+                    "with tool-call support; the current model/template does not advertise tools."
+                ),
+                status = 400,
+                code = "unsupported_parameter",
+                param = "tools" if payload.tools else "messages",
+            ),
+        )
+    if (
+        using_gguf
+        and not _studio_tool_loop_requested
         and (_tools_passthrough or _has_response_format)
     ):
         if _wants_multiple_choices(payload):
@@ -6104,12 +6736,20 @@ async def openai_chat_completions(
                 completion_id,
                 monitor_id = monitor_id,
             )
-        return await _openai_passthrough_non_streaming(
-            llama_backend,
-            payload,
-            model_name,
-            monitor_id = monitor_id,
-        )
+        _cancel_keys = (payload.cancel_id, payload.session_id, completion_id)
+        _tracker = _TrackedCancel(cancel_event, *_cancel_keys)
+        _tracker.__enter__()
+        try:
+            return await _openai_passthrough_non_streaming(
+                llama_backend,
+                payload,
+                model_name,
+                monitor_id = monitor_id,
+                request = request,
+                cancel_event = cancel_event,
+            )
+        finally:
+            _tracker.__exit__(None, None, None)
 
     # ── Parse messages (handles multimodal content parts) ─────
     # Reuse the pre-hook parse when auto-switch did it, else parse now.
@@ -6169,6 +6809,8 @@ async def openai_chat_completions(
             )
 
         def _gguf_chat_delta_line(delta: ChoiceDelta, finish_reason = None) -> str:
+            if delta.reasoning_content is not None and delta.content is None:
+                delta = delta.model_copy(update = {"content": ""})
             chunk = ChatCompletionChunk(
                 id = completion_id,
                 created = created,
@@ -6192,8 +6834,12 @@ async def openai_chat_completions(
         from state.tool_policy import get_tool_policy as _get_tool_policy_g
 
         _cli_policy = _get_tool_policy_g()
-        _tools_on = _effective_enable_tools(payload)
-        _mcp_allowed = bool(payload.mcp_enabled) and _cli_policy is not False
+        _tools_on = False if _client_disabled_tool_calls else _effective_enable_tools(payload)
+        _mcp_allowed = (
+            not _client_disabled_tool_calls
+            and bool(payload.mcp_enabled)
+            and _cli_policy is not False
+        )
         use_tools = (_tools_on or _mcp_allowed) and llama_backend.supports_tools
 
         if use_tools:
@@ -6292,6 +6938,24 @@ async def openai_chat_completions(
                     bypass_permissions = bool(payload.bypass_permissions),
                 )
 
+            _tool_admission_mode = "chat_tool_stream" if payload.stream else "chat_tool_nonstream"
+            try:
+                reservation, admission_config = _openai_llama_admission_reserve(
+                    request = request,
+                    llama_backend = llama_backend,
+                )
+            except LlamaAdmissionQueueFull as exc:
+                _openai_admission_log(
+                    "queue-full",
+                    snapshot = exc.snapshot,
+                    request = request,
+                    mode = _tool_admission_mode,
+                    completion_id = completion_id,
+                    level = "warning",
+                )
+                api_monitor.fail(monitor_id, str(exc))
+                raise _openai_admission_http_exception(exc, status_code = 429)
+
             _tool_sentinel = object()
 
             _cancel_keys = (payload.cancel_id, payload.session_id, completion_id)
@@ -6300,6 +6964,8 @@ async def openai_chat_completions(
 
             async def gguf_tool_stream():
                 gen = None
+                next_task = None
+                stream_completed = False
                 disconnect_watcher = asyncio.create_task(
                     _await_disconnect_then_cancel(request, cancel_event)
                 )
@@ -6337,7 +7003,14 @@ async def openai_chat_completions(
                             api_monitor.finish(monitor_id, "cancelled")
                             return
 
-                        event = await asyncio.to_thread(next, gen, _tool_sentinel)
+                        next_task = asyncio.create_task(
+                            asyncio.to_thread(next, gen, _tool_sentinel)
+                        )
+                        try:
+                            event = await asyncio.shield(next_task)
+                        finally:
+                            if next_task.done():
+                                next_task = None
                         if event is _tool_sentinel:
                             break
 
@@ -6436,6 +7109,7 @@ async def openai_chat_completions(
                     api_monitor.finish(
                         monitor_id, "cancelled" if cancel_event.is_set() else "completed"
                     )
+                    stream_completed = True
                     yield "data: [DONE]\n\n"
 
                 except asyncio.CancelledError:
@@ -6448,20 +7122,157 @@ async def openai_chat_completions(
                     # Recover if an MTP+tensor crash killed the server mid-stream.
                     get_llama_cpp_backend()._maybe_recover_from_mtp_crash(e)
                     error_chunk = _openai_stream_error_chunk(e)
-                    yield f"data: {json.dumps(error_chunk)}\n\n"
+                    yield _openai_stream_error_sse(error_chunk)
                 finally:
-                    await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
-                    if gen is not None:
-                        try:
-                            gen.close()
-                        except (RuntimeError, ValueError):
-                            pass
-                    _tracker.__exit__(None, None, None)
+                    try:
+                        if not stream_completed:
+                            cancel_event.set()
+                        task_to_drain = next_task
+                        next_task = None
+                        while task_to_drain is not None and not task_to_drain.done():
+                            try:
+                                await asyncio.shield(task_to_drain)
+                            except asyncio.CancelledError:
+                                cancel_event.set()
+                                continue
+                            except Exception:
+                                break
+                        if task_to_drain is not None and task_to_drain.done():
+                            try:
+                                task_to_drain.exception()
+                            except (asyncio.CancelledError, Exception):
+                                pass
+                        if gen is not None and not stream_completed:
+                            try:
+                                await asyncio.to_thread(gen.close)
+                            except (RuntimeError, ValueError):
+                                pass
+                            except Exception:
+                                logger.debug(
+                                    "Error closing GGUF tool stream generator during cleanup",
+                                    exc_info = True,
+                                )
+                        await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
+                    finally:
+                        _tracker.__exit__(None, None, None)
 
             if payload.stream:
+                stream_lease = reservation.lease_nowait()
+                admission_wait_started_at = None
+                if stream_lease is None:
+                    admission_wait_started_at = time.monotonic()
+                    _openai_admission_log(
+                        "queued",
+                        reservation,
+                        request = request,
+                        mode = _tool_admission_mode,
+                        completion_id = completion_id,
+                        level = "debug",
+                    )
+
+                async def admitted_gguf_tool_stream():
+                    lease = stream_lease
+                    stream_started = False
+                    stream_cancelled = False
+                    try:
+                        if lease is None:
+                            async for wait_item in _openai_admission_wait_stream_chunks(
+                                reservation,
+                                admission_config,
+                                request = request,
+                                cancel_event = cancel_event,
+                            ):
+                                if isinstance(wait_item, str):
+                                    yield wait_item
+                                    continue
+                                lease = wait_item
+                                _openai_admission_log(
+                                    "granted-after-wait",
+                                    reservation,
+                                    request = request,
+                                    mode = _tool_admission_mode,
+                                    wait_started_at = admission_wait_started_at,
+                                    completion_id = completion_id,
+                                    level = "debug",
+                                )
+                                break
+                        if lease is None:
+                            return
+                        await _raise_if_openai_admission_cancelled(
+                            reservation,
+                            request = request,
+                            cancel_event = cancel_event,
+                        )
+                        iterator = gguf_tool_stream()
+                        stream_started = True
+                        try:
+                            async for chunk in iterator:
+                                yield chunk
+                        except asyncio.CancelledError:
+                            stream_cancelled = True
+                            raise
+                        finally:
+                            await _close_openai_admitted_stream_iterator(
+                                iterator,
+                                cancelled = stream_cancelled,
+                            )
+                    except LlamaAdmissionTimeout as exc:
+                        _openai_admission_log(
+                            "timeout",
+                            reservation,
+                            request = request,
+                            mode = _tool_admission_mode,
+                            wait_started_at = admission_wait_started_at,
+                            completion_id = completion_id,
+                            level = "warning",
+                        )
+                        api_monitor.fail(monitor_id, str(exc))
+                        yield _openai_stream_error_sse(
+                            _openai_admission_error_body(exc, status_code = 503)
+                        )
+                    except LlamaAdmissionCancelled:
+                        _openai_admission_log(
+                            "cancelled-before-upstream",
+                            reservation,
+                            request = request,
+                            mode = _tool_admission_mode,
+                            wait_started_at = admission_wait_started_at,
+                            completion_id = completion_id,
+                            level = "debug",
+                        )
+                        api_monitor.finish(monitor_id, "cancelled")
+                        return
+                    except asyncio.CancelledError:
+                        api_monitor.finish(monitor_id, "cancelled")
+                        raise
+                    except HTTPException as exc:
+                        status_code = getattr(exc, "status_code", 500) or 500
+                        detail = exc.detail
+                        error = (
+                            detail
+                            if isinstance(detail, dict) and "error" in detail
+                            else openai_error_body(str(detail), status = status_code)
+                        )
+                        api_monitor.fail(monitor_id, str(detail))
+                        yield _openai_stream_error_sse(error)
+                    finally:
+                        if lease is not None:
+                            lease.release()
+                        if not stream_started:
+                            api_monitor.finish(monitor_id, "cancelled")
+                            reservation.cancel()
+                            _tracker.__exit__(None, None, None)
+
+                async def _gguf_tool_admission_unstarted_cleanup() -> None:
+                    api_monitor.finish(monitor_id, "cancelled")
+                    if stream_lease is not None:
+                        stream_lease.release()
+                    reservation.cancel()
+                    _tracker.__exit__(None, None, None)
+
                 return _SameTaskStreamingResponse(
-                    gguf_tool_stream(),
-                    unstarted_cleanup = _tracked_cancel_unstarted_cleanup(_tracker),
+                    admitted_gguf_tool_stream(),
+                    unstarted_cleanup = _gguf_tool_admission_unstarted_cleanup,
                     media_type = "text/event-stream",
                     headers = {
                         "Cache-Control": "no-cache",
@@ -6507,10 +7318,61 @@ async def openai_chat_completions(
                     except (RuntimeError, ValueError):
                         pass
 
+            drain_task = None
+
+            async def _drain_cancelled_gguf_tool_task():
+                if drain_task is None:
+                    return
+                while not drain_task.done():
+                    try:
+                        await asyncio.shield(drain_task)
+                    except asyncio.CancelledError:
+                        cancel_event.set()
+                        continue
+                    except Exception:
+                        break
+                if drain_task.done():
+                    try:
+                        drain_task.exception()
+                    except (asyncio.CancelledError, Exception):
+                        pass
+
+            admission_lease = None
+            admission_wait_started_at = None
             try:
-                full_text, completion_usage, completion_finish = await asyncio.to_thread(
-                    _drain_gguf_tool_loop
+                if reservation.lease_nowait() is None:
+                    admission_wait_started_at = time.monotonic()
+                    _openai_admission_log(
+                        "queued",
+                        reservation,
+                        request = request,
+                        mode = _tool_admission_mode,
+                        completion_id = completion_id,
+                        level = "debug",
+                    )
+                admission_lease = await _wait_for_openai_admission_non_streaming(
+                    reservation,
+                    admission_config,
+                    request = request,
+                    cancel_event = cancel_event,
                 )
+                if admission_wait_started_at is not None:
+                    _openai_admission_log(
+                        "granted-after-wait",
+                        reservation,
+                        request = request,
+                        mode = _tool_admission_mode,
+                        wait_started_at = admission_wait_started_at,
+                        completion_id = completion_id,
+                        level = "debug",
+                    )
+                await _raise_if_openai_admission_cancelled(
+                    reservation,
+                    request = request,
+                    cancel_event = cancel_event,
+                )
+                drain_task = asyncio.create_task(asyncio.to_thread(_drain_gguf_tool_loop))
+                full_text, completion_usage, completion_finish = await asyncio.shield(drain_task)
                 reasoning_text, visible_text = _extract_responses_reasoning(
                     full_text,
                     parse_think_markers = _responses_should_parse_think_markers(
@@ -6556,6 +7418,48 @@ async def openai_chat_completions(
                     monitor_id, "cancelled" if cancel_event.is_set() else "completed"
                 )
                 return _model_json_response(response)
+            except asyncio.CancelledError:
+                cancel_event.set()
+                await _drain_cancelled_gguf_tool_task()
+                api_monitor.finish(monitor_id, "cancelled")
+                reservation.cancel()
+                if admission_lease is not None:
+                    admission_lease.release()
+                _tracker.__exit__(None, None, None)
+                raise
+            except LlamaAdmissionTimeout as exc:
+                _openai_admission_log(
+                    "timeout",
+                    reservation,
+                    request = request,
+                    mode = _tool_admission_mode,
+                    wait_started_at = admission_wait_started_at,
+                    completion_id = completion_id,
+                    level = "warning",
+                )
+                api_monitor.fail(monitor_id, str(exc))
+                if admission_lease is not None:
+                    admission_lease.release()
+                _tracker.__exit__(None, None, None)
+                raise _openai_admission_http_exception(exc, status_code = 503)
+            except LlamaAdmissionCancelled as exc:
+                _openai_admission_log(
+                    "cancelled-before-upstream",
+                    reservation,
+                    request = request,
+                    mode = _tool_admission_mode,
+                    wait_started_at = admission_wait_started_at,
+                    completion_id = completion_id,
+                    level = "debug",
+                )
+                api_monitor.finish(monitor_id, "cancelled")
+                if admission_lease is not None:
+                    admission_lease.release()
+                _tracker.__exit__(None, None, None)
+                raise HTTPException(
+                    status_code = 499,
+                    detail = _openai_admission_error_body(exc, status_code = 499),
+                )
             except Exception as e:
                 logger.error(f"Error during GGUF tool completion: {e}", exc_info = True)
                 api_monitor.fail(monitor_id, _friendly_error(e))
@@ -6576,6 +7480,8 @@ async def openai_chat_completions(
                     )
                 raise HTTPException(status_code = 500, detail = safe_error_detail(e))
             finally:
+                if admission_lease is not None:
+                    admission_lease.release()
                 _tracker.__exit__(None, None, None)
 
         # ── Standard GGUF path (no tools) ─────────────────────
@@ -6610,11 +7516,31 @@ async def openai_chat_completions(
             _cancel_keys = (payload.cancel_id, payload.session_id, completion_id)
             _tracker = _TrackedCancel(cancel_event, *_cancel_keys)
             _tracker.__enter__()
+            try:
+                reservation, admission_config = _openai_llama_admission_reserve(
+                    request = request,
+                    llama_backend = llama_backend,
+                )
+            except LlamaAdmissionQueueFull as exc:
+                _tracker.__exit__(None, None, None)
+                _openai_admission_log(
+                    "queue-full",
+                    snapshot = exc.snapshot,
+                    request = request,
+                    mode = "chat_standard_stream",
+                    completion_id = completion_id,
+                    level = "warning",
+                )
+                api_monitor.fail(monitor_id, str(exc))
+                raise _openai_admission_http_exception(exc, status_code = 429)
 
             async def gguf_stream_chunks():
                 disconnect_watcher = asyncio.create_task(
                     _await_disconnect_then_cancel(request, cancel_event)
                 )
+                gen = None
+                next_task = None
+                stream_completed = False
                 try:
                     yield _chat_role_chunk(completion_id, created, model_name)
 
@@ -6633,7 +7559,14 @@ async def openai_chat_completions(
                             cancel_event.set()
                             api_monitor.finish(monitor_id, "cancelled")
                             return
-                        cumulative = await asyncio.to_thread(next, gen, _gguf_sentinel)
+                        next_task = asyncio.create_task(
+                            asyncio.to_thread(next, gen, _gguf_sentinel)
+                        )
+                        try:
+                            cumulative = await asyncio.shield(next_task)
+                        finally:
+                            if next_task.done():
+                                next_task = None
                         if cumulative is _gguf_sentinel:
                             break
                         # Capture server metadata for the final usage chunk
@@ -6702,6 +7635,7 @@ async def openai_chat_completions(
                     api_monitor.finish(
                         monitor_id, "cancelled" if cancel_event.is_set() else "completed"
                     )
+                    stream_completed = True
                     yield "data: [DONE]\n\n"
 
                 except asyncio.CancelledError:
@@ -6712,14 +7646,156 @@ async def openai_chat_completions(
                     logger.error(f"Error during GGUF streaming: {e}", exc_info = True)
                     api_monitor.fail(monitor_id, _friendly_error(e))
                     error_chunk = _openai_stream_error_chunk(e)
-                    yield f"data: {json.dumps(error_chunk)}\n\n"
+                    yield _openai_stream_error_sse(error_chunk)
                 finally:
-                    await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
-                    _tracker.__exit__(None, None, None)
+                    try:
+                        if not stream_completed:
+                            cancel_event.set()
+                        task_to_drain = next_task
+                        next_task = None
+                        while task_to_drain is not None and not task_to_drain.done():
+                            try:
+                                await asyncio.shield(task_to_drain)
+                            except asyncio.CancelledError:
+                                cancel_event.set()
+                                continue
+                            except Exception:
+                                break
+                        if task_to_drain is not None and task_to_drain.done():
+                            try:
+                                task_to_drain.exception()
+                            except (asyncio.CancelledError, Exception):
+                                pass
+                        if gen is not None and not stream_completed:
+                            try:
+                                await asyncio.to_thread(gen.close)
+                            except (RuntimeError, ValueError):
+                                pass
+                            except Exception:
+                                logger.debug(
+                                    "Error closing GGUF stream generator during cleanup",
+                                    exc_info = True,
+                                )
+                        await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
+                    finally:
+                        _tracker.__exit__(None, None, None)
+
+            stream_lease = reservation.lease_nowait()
+            admission_wait_started_at = None
+            if stream_lease is None:
+                admission_wait_started_at = time.monotonic()
+                _openai_admission_log(
+                    "queued",
+                    reservation,
+                    request = request,
+                    mode = "chat_standard_stream",
+                    completion_id = completion_id,
+                    level = "debug",
+                )
+
+            async def admitted_gguf_stream_chunks():
+                lease = stream_lease
+                stream_started = False
+                stream_cancelled = False
+                try:
+                    if lease is None:
+                        async for wait_item in _openai_admission_wait_stream_chunks(
+                            reservation,
+                            admission_config,
+                            request = request,
+                            cancel_event = cancel_event,
+                        ):
+                            if isinstance(wait_item, str):
+                                yield wait_item
+                                continue
+                            lease = wait_item
+                            _openai_admission_log(
+                                "granted-after-wait",
+                                reservation,
+                                request = request,
+                                mode = "chat_standard_stream",
+                                wait_started_at = admission_wait_started_at,
+                                completion_id = completion_id,
+                                level = "debug",
+                            )
+                            break
+                    if lease is None:
+                        return
+                    await _raise_if_openai_admission_cancelled(
+                        reservation,
+                        request = request,
+                        cancel_event = cancel_event,
+                    )
+                    iterator = gguf_stream_chunks()
+                    stream_started = True
+                    try:
+                        async for chunk in iterator:
+                            yield chunk
+                    except asyncio.CancelledError:
+                        stream_cancelled = True
+                        raise
+                    finally:
+                        await _close_openai_admitted_stream_iterator(
+                            iterator,
+                            cancelled = stream_cancelled,
+                        )
+                except LlamaAdmissionTimeout as exc:
+                    _openai_admission_log(
+                        "timeout",
+                        reservation,
+                        request = request,
+                        mode = "chat_standard_stream",
+                        wait_started_at = admission_wait_started_at,
+                        completion_id = completion_id,
+                        level = "warning",
+                    )
+                    api_monitor.fail(monitor_id, str(exc))
+                    yield _openai_stream_error_sse(
+                        _openai_admission_error_body(exc, status_code = 503)
+                    )
+                except LlamaAdmissionCancelled:
+                    _openai_admission_log(
+                        "cancelled-before-upstream",
+                        reservation,
+                        request = request,
+                        mode = "chat_standard_stream",
+                        wait_started_at = admission_wait_started_at,
+                        completion_id = completion_id,
+                        level = "debug",
+                    )
+                    api_monitor.finish(monitor_id, "cancelled")
+                    return
+                except asyncio.CancelledError:
+                    api_monitor.finish(monitor_id, "cancelled")
+                    raise
+                except HTTPException as exc:
+                    status_code = getattr(exc, "status_code", 500) or 500
+                    detail = exc.detail
+                    error = (
+                        detail
+                        if isinstance(detail, dict) and "error" in detail
+                        else openai_error_body(str(detail), status = status_code)
+                    )
+                    api_monitor.fail(monitor_id, str(detail))
+                    yield _openai_stream_error_sse(error)
+                finally:
+                    if lease is not None:
+                        lease.release()
+                    if not stream_started:
+                        api_monitor.finish(monitor_id, "cancelled")
+                        reservation.cancel()
+                        _tracker.__exit__(None, None, None)
+
+            async def _gguf_admission_unstarted_cleanup() -> None:
+                api_monitor.finish(monitor_id, "cancelled")
+                if stream_lease is not None:
+                    stream_lease.release()
+                reservation.cancel()
+                _tracker.__exit__(None, None, None)
 
             return _SameTaskStreamingResponse(
-                gguf_stream_chunks(),
-                unstarted_cleanup = _tracked_cancel_unstarted_cleanup(_tracker),
+                admitted_gguf_stream_chunks(),
+                unstarted_cleanup = _gguf_admission_unstarted_cleanup,
                 media_type = "text/event-stream",
                 headers = {
                     "Cache-Control": "no-cache",
@@ -6728,57 +7804,189 @@ async def openai_chat_completions(
                 },
             )
         else:
+            try:
+                reservation, admission_config = _openai_llama_admission_reserve(
+                    request = request,
+                    llama_backend = llama_backend,
+                )
+            except LlamaAdmissionQueueFull as exc:
+                _openai_admission_log(
+                    "queue-full",
+                    snapshot = exc.snapshot,
+                    request = request,
+                    mode = "chat_standard_nonstream",
+                    completion_id = completion_id,
+                    level = "warning",
+                )
+                api_monitor.fail(monitor_id, str(exc))
+                raise _openai_admission_http_exception(exc, status_code = 429)
+
+            _cancel_keys = (payload.cancel_id, payload.session_id, completion_id)
+            _tracker = _TrackedCancel(cancel_event, *_cancel_keys)
+            _tracker.__enter__()
+            admission_lease = None
+            admission_wait_started_at = None
+            try:
+                if reservation.lease_nowait() is None:
+                    admission_wait_started_at = time.monotonic()
+                    _openai_admission_log(
+                        "queued",
+                        reservation,
+                        request = request,
+                        mode = "chat_standard_nonstream",
+                        completion_id = completion_id,
+                        level = "debug",
+                    )
+                admission_lease = await _wait_for_openai_admission_non_streaming(
+                    reservation,
+                    admission_config,
+                    request = request,
+                    cancel_event = cancel_event,
+                )
+                if admission_wait_started_at is not None:
+                    _openai_admission_log(
+                        "granted-after-wait",
+                        reservation,
+                        request = request,
+                        mode = "chat_standard_nonstream",
+                        wait_started_at = admission_wait_started_at,
+                        completion_id = completion_id,
+                        level = "debug",
+                    )
+                await _raise_if_openai_admission_cancelled(
+                    reservation,
+                    request = request,
+                    cancel_event = cancel_event,
+                )
+            except asyncio.CancelledError:
+                api_monitor.finish(monitor_id, "cancelled")
+                reservation.cancel()
+                if admission_lease is not None:
+                    admission_lease.release()
+                _tracker.__exit__(None, None, None)
+                raise
+            except LlamaAdmissionTimeout as exc:
+                _openai_admission_log(
+                    "timeout",
+                    reservation,
+                    request = request,
+                    mode = "chat_standard_nonstream",
+                    wait_started_at = admission_wait_started_at,
+                    completion_id = completion_id,
+                    level = "warning",
+                )
+                api_monitor.fail(monitor_id, str(exc))
+                if admission_lease is not None:
+                    admission_lease.release()
+                _tracker.__exit__(None, None, None)
+                raise _openai_admission_http_exception(exc, status_code = 503)
+            except LlamaAdmissionCancelled as exc:
+                _openai_admission_log(
+                    "cancelled-before-upstream",
+                    reservation,
+                    request = request,
+                    mode = "chat_standard_nonstream",
+                    wait_started_at = admission_wait_started_at,
+                    completion_id = completion_id,
+                    level = "debug",
+                )
+                api_monitor.finish(monitor_id, "cancelled")
+                if admission_lease is not None:
+                    admission_lease.release()
+                _tracker.__exit__(None, None, None)
+                raise HTTPException(
+                    status_code = 499,
+                    detail = _openai_admission_error_body(exc, status_code = 499),
+                )
+
             try:
                 # ``n`` requests several independent completions; the single
                 # decode slot yields one at a time, so loop sequentially.
-                _n = payload.n or 1
+                drain_task = None
 
-                _choices = []
-                _monitor_replies = []
-                _prompt_tokens = 0
-                _sum_completion = 0
-                _prompt_details = None
-                for _idx in range(_n):
-                    # Stop spawning the remaining choices once cancelled.
-                    if cancel_event.is_set():
-                        break
-                    full_text = ""
-                    completion_usage = None
-                    completion_finish = None
-                    for token in gguf_generate(_idx):
-                        if isinstance(token, dict):
-                            if token.get("type") == "metadata":
-                                completion_usage = token.get("usage")
-                                completion_finish = token.get("finish_reason")
+                async def _drain_cancelled_gguf_task():
+                    if drain_task is None:
+                        return
+                    while not drain_task.done():
+                        try:
+                            await asyncio.shield(drain_task)
+                        except asyncio.CancelledError:
+                            cancel_event.set()
                             continue
-                        full_text = token
+                        except Exception:
+                            break
+                    if drain_task.done():
+                        try:
+                            drain_task.exception()
+                        except (asyncio.CancelledError, Exception):
+                            pass
 
-                    reasoning_text, visible_text = _extract_responses_reasoning(
-                        full_text,
-                        parse_think_markers = _responses_should_parse_think_markers(
-                            payload,
-                            llama_backend,
-                        ),
-                    )
-                    message_kwargs = {"content": visible_text}
-                    if reasoning_text:
-                        message_kwargs["reasoning_content"] = reasoning_text
-                    _choices.append(
-                        CompletionChoice(
-                            index = _idx,
-                            message = CompletionMessage(**message_kwargs),
-                            finish_reason = _clamp_finish_reason(completion_finish),
+                def _drain_gguf_choices():
+                    _n = payload.n or 1
+                    _choices = []
+                    _monitor_replies = []
+                    _prompt_tokens = 0
+                    _sum_completion = 0
+                    _prompt_details = None
+                    for _idx in range(_n):
+                        # Stop spawning the remaining choices once cancelled.
+                        if cancel_event.is_set():
+                            break
+                        full_text = ""
+                        completion_usage = None
+                        completion_finish = None
+                        for token in gguf_generate(_idx):
+                            if isinstance(token, dict):
+                                if token.get("type") == "metadata":
+                                    completion_usage = token.get("usage")
+                                    completion_finish = token.get("finish_reason")
+                                continue
+                            full_text = token
+
+                        reasoning_text, visible_text = _extract_responses_reasoning(
+                            full_text,
+                            parse_think_markers = _responses_should_parse_think_markers(
+                                payload,
+                                llama_backend,
+                            ),
                         )
+                        message_kwargs = {"content": visible_text}
+                        if reasoning_text:
+                            message_kwargs["reasoning_content"] = reasoning_text
+                        _choices.append(
+                            CompletionChoice(
+                                index = _idx,
+                                message = CompletionMessage(**message_kwargs),
+                                finish_reason = _clamp_finish_reason(completion_finish),
+                            )
+                        )
+                        _monitor_replies.append(visible_text)
+                        if completion_usage:
+                            # The prompt is shared across all n choices, so count its
+                            # tokens ONCE (OpenAI bills only generated tokens for each
+                            # extra choice). Only completion_tokens accumulates.
+                            _prompt_tokens = completion_usage.get("prompt_tokens") or _prompt_tokens
+                            _sum_completion += completion_usage.get("completion_tokens") or 0
+                            if _prompt_details is None:
+                                _prompt_details = completion_usage.get("prompt_tokens_details")
+                    return (
+                        _n,
+                        _choices,
+                        _monitor_replies,
+                        _prompt_tokens,
+                        _sum_completion,
+                        _prompt_details,
                     )
-                    _monitor_replies.append(visible_text)
-                    if completion_usage:
-                        # The prompt is shared across all n choices, so count its
-                        # tokens ONCE (OpenAI bills only generated tokens for each
-                        # extra choice). Only completion_tokens accumulates.
-                        _prompt_tokens = completion_usage.get("prompt_tokens") or _prompt_tokens
-                        _sum_completion += completion_usage.get("completion_tokens") or 0
-                        if _prompt_details is None:
-                            _prompt_details = completion_usage.get("prompt_tokens_details")
+
+                drain_task = asyncio.create_task(asyncio.to_thread(_drain_gguf_choices))
+                (
+                    _n,
+                    _choices,
+                    _monitor_replies,
+                    _prompt_tokens,
+                    _sum_completion,
+                    _prompt_details,
+                ) = await asyncio.shield(drain_task)
 
                 response = ChatCompletion(
                     id = completion_id,
@@ -6810,6 +8018,11 @@ async def openai_chat_completions(
                 api_monitor.finish(monitor_id)
                 return _model_json_response(response)
 
+            except asyncio.CancelledError:
+                cancel_event.set()
+                await _drain_cancelled_gguf_task()
+                api_monitor.finish(monitor_id, "cancelled")
+                raise
             except Exception as e:
                 logger.error(f"Error during GGUF completion: {e}", exc_info = True)
                 api_monitor.fail(monitor_id, _friendly_error(e))
@@ -6829,7 +8042,10 @@ async def openai_chat_completions(
                         ),
                     )
                 raise HTTPException(status_code = 500, detail = safe_error_detail(e))
-
+            finally:
+                if admission_lease is not None:
+                    admission_lease.release()
+                _tracker.__exit__(None, None, None)
     # ── Standard Unsloth path ─────────────────────────────────
 
     # Decode image (from content parts OR legacy field)
@@ -7149,7 +8365,7 @@ async def openai_chat_completions(
                         "type": "server_error",
                     },
                 }
-                yield f"data: {json.dumps(error_chunk)}\n\n"
+                yield _openai_stream_error_sse(error_chunk)
             finally:
                 await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
                 if gen is not None:
@@ -7494,7 +8710,7 @@ async def openai_chat_completions(
                         "type": "server_error",
                     },
                 }
-                yield f"data: {json.dumps(error_chunk)}\n\n"
+                yield _openai_stream_error_sse(error_chunk)
             finally:
                 await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
                 _tracker.__exit__(None, None, None)
@@ -7983,7 +9199,7 @@ async def openai_completions(request: Request, current_subject: str = Depends(ge
     if not llama_backend.is_loaded:
         raise HTTPException(
             status_code = 503,
-            detail = "No GGUF model loaded. Load a GGUF model first.",
+            detail = _no_model_loaded_detail("No GGUF model loaded. Load a GGUF model first."),
         )
     if not isinstance(body, dict):
         # Re-read to re-raise a malformed-body error (post-503, pre-feature behavior);
@@ -7992,8 +9208,12 @@ async def openai_completions(request: Request, current_subject: str = Depends(ge
         if not isinstance(body, dict):
             raise HTTPException(status_code = 400, detail = "Request body must be a JSON object")
 
-    if body.get("max_tokens") is None:
-        body["max_tokens"] = llama_backend.context_length or _DEFAULT_MAX_TOKENS_FLOOR
+    _resolved_max_tokens = _effective_openai_max_tokens_from_values(body.get("max_tokens"))
+    body["max_tokens"] = (
+        _resolved_max_tokens
+        if _resolved_max_tokens is not None
+        else (llama_backend.context_length or _DEFAULT_MAX_TOKENS_FLOOR)
+    )
     target_url = f"{llama_backend.base_url}/v1/completions"
     is_stream = body.get("stream", False)
     prompt_text = _flatten_monitor_prompt(body.get("prompt", ""))
@@ -8086,7 +9306,7 @@ async def openai_completions(request: Request, current_subject: str = Depends(ge
                     logger.error("openai_completions stream error: %s", e)
                     api_monitor.fail(monitor_id, _friendly_error(e))
                     error_chunk = _openai_stream_error_chunk(e)
-                    yield f"data: {json.dumps(error_chunk)}\n\n".encode("utf-8")
+                    yield _openai_stream_error_sse_bytes(error_chunk)
                     return
                 api_monitor.finish(monitor_id, "cancelled")
                 return
@@ -8101,7 +9321,7 @@ async def openai_completions(request: Request, current_subject: str = Depends(ge
                 logger.error("openai_completions stream error: %s", e)
                 api_monitor.fail(monitor_id, _friendly_error(e))
                 error_chunk = _openai_stream_error_chunk(e)
-                yield f"data: {json.dumps(error_chunk)}\n\n".encode("utf-8")
+                yield _openai_stream_error_sse_bytes(error_chunk)
                 return
             finally:
                 await _aclose_stream_resources(
@@ -8195,7 +9415,7 @@ async def openai_embeddings(request: Request, current_subject: str = Depends(get
     if not llama_backend.is_loaded:
         raise HTTPException(
             status_code = 503,
-            detail = "No GGUF model loaded. Load a GGUF model first.",
+            detail = _no_model_loaded_detail("No GGUF model loaded. Load a GGUF model first."),
         )
     if not isinstance(body, dict):
         # Re-read to re-raise a malformed-body error (post-503, pre-feature behavior);
@@ -8935,7 +10155,7 @@ async def _responses_stream(
         # so the client sees a useful error instead of a dangling stream.
         raise HTTPException(
             status_code = 400,
-            detail = (
+            detail = _no_model_loaded_detail(
                 "Streaming /v1/responses requires a GGUF model loaded via "
                 "llama-server. Use non-streaming /v1/responses, "
                 "/v1/chat/completions, or load a GGUF model."
@@ -8957,6 +10177,52 @@ async def _responses_stream(
     )
     body["stream_options"] = {"include_usage": True}
     target_url = f"{llama_backend.base_url}/v1/chat/completions"
+    try:
+        reservation, admission_config = _openai_llama_admission_reserve(
+            request = request,
+            llama_backend = llama_backend,
+        )
+    except LlamaAdmissionQueueFull as exc:
+        _openai_admission_log(
+            "queue-full",
+            snapshot = exc.snapshot,
+            request = request,
+            mode = "responses_stream",
+            completion_id = resp_id,
+            level = "warning",
+        )
+        api_monitor.fail(monitor_id, str(exc))
+        raise _openai_admission_http_exception(exc, status_code = 429)
+
+    def _responses_admission_failed_sse(exc: Exception, *, status_code: int) -> str:
+        return (
+            "event: response.failed\n"
+            "data: "
+            + json.dumps(
+                {
+                    "type": "response.failed",
+                    "response": {
+                        "id": resp_id,
+                        "object": "response",
+                        "created_at": created_at,
+                        "status": "failed",
+                        "model": _llama_public_model_id(llama_backend, payload.model)
+                        or payload.model,
+                        "output": [],
+                        "usage": {
+                            "input_tokens": 0,
+                            "output_tokens": 0,
+                            "total_tokens": 0,
+                        },
+                        "error": {
+                            "code": status_code,
+                            "message": str(exc),
+                        },
+                    },
+                }
+            )
+            + "\n\n"
+        )
 
     async def event_generator():
         # Clean public id for every response envelope. Prefer the loaded model's
@@ -9766,14 +11032,111 @@ async def _responses_stream(
         api_monitor.finish(monitor_id)
         yield _sse("response.completed", completed_response)
 
+    async def admitted_event_generator():
+        lease = reservation.lease_nowait()
+        admission_wait_started_at = None
+        stream_started = False
+        stream_cancelled = False
+        iterator = None
+        try:
+            if lease is None:
+                admission_wait_started_at = time.monotonic()
+                _openai_admission_log(
+                    "queued",
+                    reservation,
+                    request = request,
+                    mode = "responses_stream",
+                    completion_id = resp_id,
+                    level = "debug",
+                )
+                async for wait_item in _openai_admission_wait_stream_chunks(
+                    reservation,
+                    admission_config,
+                    request = request,
+                    cancel_event = None,
+                ):
+                    if isinstance(wait_item, str):
+                        yield wait_item
+                        continue
+                    lease = wait_item
+                    _openai_admission_log(
+                        "granted-after-wait",
+                        reservation,
+                        request = request,
+                        mode = "responses_stream",
+                        wait_started_at = admission_wait_started_at,
+                        completion_id = resp_id,
+                        level = "debug",
+                    )
+                    break
+            if lease is None:
+                return
+            await _raise_if_openai_admission_cancelled(
+                reservation,
+                request = request,
+                cancel_event = None,
+            )
+            iterator = event_generator()
+            stream_started = True
+            try:
+                async for chunk in iterator:
+                    yield chunk
+            except asyncio.CancelledError:
+                stream_cancelled = True
+                api_monitor.finish(monitor_id, "cancelled")
+                raise
+            finally:
+                await _close_openai_admitted_stream_iterator(
+                    iterator,
+                    cancelled = stream_cancelled,
+                )
+        except LlamaAdmissionTimeout as exc:
+            _openai_admission_log(
+                "timeout",
+                reservation,
+                request = request,
+                mode = "responses_stream",
+                wait_started_at = admission_wait_started_at,
+                completion_id = resp_id,
+                level = "warning",
+            )
+            api_monitor.fail(monitor_id, str(exc))
+            yield _responses_admission_failed_sse(exc, status_code = 503)
+        except LlamaAdmissionCancelled:
+            _openai_admission_log(
+                "cancelled-before-upstream",
+                reservation,
+                request = request,
+                mode = "responses_stream",
+                wait_started_at = admission_wait_started_at,
+                completion_id = resp_id,
+                level = "debug",
+            )
+            api_monitor.finish(monitor_id, "cancelled")
+            return
+        except asyncio.CancelledError:
+            api_monitor.finish(monitor_id, "cancelled")
+            raise
+        finally:
+            if lease is not None:
+                lease.release()
+            if not stream_started:
+                api_monitor.finish(monitor_id, "cancelled")
+                reservation.cancel()
+
+    async def _responses_admission_unstarted_cleanup() -> None:
+        api_monitor.finish(monitor_id, "cancelled")
+        reservation.cancel()
+
     return _SameTaskStreamingResponse(
-        event_generator(),
+        admitted_event_generator(),
         media_type = "text/event-stream",
         headers = {
             "Cache-Control": "no-cache",
             "Connection": "close",
             "X-Accel-Buffering": "no",
         },
+        unstarted_cleanup = _responses_admission_unstarted_cleanup,
     )
 
 
@@ -10031,7 +11394,7 @@ async def anthropic_count_tokens(
     if not llama_backend.is_loaded:
         raise HTTPException(
             status_code = 503,
-            detail = "No GGUF model loaded. Load a GGUF model first.",
+            detail = _no_model_loaded_detail("No GGUF model loaded. Load a GGUF model first."),
         )
 
     # Same Anthropic → OpenAI translation as anthropic_messages: system is
@@ -10043,8 +11406,11 @@ async def anthropic_count_tokens(
     # Apply the same sanitization /messages does before generation, so the count
     # matches the prompt the real request would build (otherwise empty-assistant
     # sentinels / synthetic tool history inflate the count or hit the fallback).
-    openai_messages = _strip_provider_synthetic_tool_history(
-        _drop_empty_assistant_sentinels(openai_messages)
+    # Coalesce adjacent user turns left behind by dropping an empty / null assistant
+    # turn, so a strict GGUF chat template does not 400 on non-alternating roles
+    # (mirrors the GGUF chat path); a no-op for already-alternating histories.
+    openai_messages = _coalesce_consecutive_user_turns(
+        _strip_provider_synthetic_tool_history(_drop_empty_assistant_sentinels(openai_messages))
     )
     openai_tools = anthropic_tools_to_openai(payload.tools or []) or None
 
@@ -10101,7 +11467,7 @@ async def anthropic_messages(
     if not llama_backend.is_loaded and not _automatic_model_load_may_run():
         raise HTTPException(
             status_code = 503,
-            detail = "No GGUF model loaded. Load a GGUF model first.",
+            detail = _no_model_loaded_detail("No GGUF model loaded. Load a GGUF model first."),
         )
 
     # max_tokens is a required field on the Anthropic Messages API; real Anthropic
@@ -10152,7 +11518,7 @@ async def anthropic_messages(
     if not llama_backend.is_loaded:
         raise HTTPException(
             status_code = 503,
-            detail = "No GGUF model loaded. Load a GGUF model first.",
+            detail = _no_model_loaded_detail("No GGUF model loaded. Load a GGUF model first."),
         )
 
     # Advertised repo id after an auto-switch load, else a clean public id, never
@@ -10172,8 +11538,11 @@ async def anthropic_messages(
     # builders apply the same strip; without it an Anthropic /v1/messages caller
     # replaying a prior provider-side tool_use forwards fake builtin tool
     # history to a backend with no matching function declarations.
-    openai_messages = _strip_provider_synthetic_tool_history(
-        _drop_empty_assistant_sentinels(openai_messages)
+    # Coalesce adjacent user turns left behind by dropping an empty / null assistant
+    # turn, so a strict GGUF chat template does not 400 on non-alternating roles
+    # (mirrors the GGUF chat path); a no-op for already-alternating histories.
+    openai_messages = _coalesce_consecutive_user_turns(
+        _strip_provider_synthetic_tool_history(_drop_empty_assistant_sentinels(openai_messages))
     )
 
     # Enforce vision guard + re-encode embedded images to PNG so the Anthropic
@@ -10222,7 +11591,9 @@ async def anthropic_messages(
         and not _has_image
     )
     client_tools = (
-        not server_tools and len(openai_client_tools) > 0 and llama_backend.supports_tools
+        not server_tools
+        and len(openai_client_tools) > 0
+        and getattr(llama_backend, "supports_tool_passthrough", llama_backend.supports_tools)
     )
 
     # Anthropic tool_choice.disable_parallel_tool_use caps the response to a
@@ -10840,13 +12211,15 @@ def _build_passthrough_payload(
 ):
     body = {
         "messages": openai_messages,
-        "tools": openai_tools,
-        "tool_choice": tool_choice,
         "temperature": temperature,
         "top_p": top_p,
         "top_k": top_k,
         "stream": stream,
     }
+    if openai_tools:
+        body["tools"] = openai_tools
+        if tool_choice is not None:
+            body["tool_choice"] = tool_choice
     if seed is not None:
         body["seed"] = seed
     if stream and stream_options is not None:
@@ -11053,13 +12426,17 @@ async def _anthropic_passthrough_stream(
                     yield event
                 return
         finally:
-            await _aclose_stream_resources(
-                watchers = (cancel_watcher, disconnect_watcher),
-                iterator = lines_iter,
-                resp = resp,
-                client = client,
-            )
-            _tracker.__exit__(None, None, None)
+            # Same shape as the OpenAI passthrough: a close-time CancelledError
+            # re-raised by _aclose_stream_resources must not skip the tracker exit.
+            try:
+                await _aclose_stream_resources(
+                    watchers = (cancel_watcher, disconnect_watcher),
+                    iterator = lines_iter,
+                    resp = resp,
+                    client = client,
+                )
+            finally:
+                _tracker.__exit__(None, None, None)
 
         for line in emitter.finish():
             yield line
@@ -11556,6 +12933,9 @@ def _build_openai_passthrough_body(
     system_prompt, _, _ = _extract_content_parts(payload.messages)
     messages = _set_or_prepend_system_message(messages, system_prompt)
     tool_choice = payload.tool_choice if payload.tool_choice is not None else "auto"
+    tools = payload.tools
+    if payload.tool_choice == "none" and not _has_openai_tool_history(payload.messages):
+        tools = None
     # Forward per-request reasoning fields (enable_thinking / reasoning_effort /
     # preserve_thinking) via chat_template_kwargs so the Jinja template renders
     # in the caller's mode, gated on the active template's capabilities exactly
@@ -11571,12 +12951,12 @@ def _build_openai_passthrough_body(
     )
     return _build_passthrough_payload(
         messages,
-        payload.tools,
+        tools,
         payload.temperature,
         payload.top_p,
         payload.top_k,
         # Honor max_completion_tokens on the tools/response_format passthrough too.
-        _effective_max_tokens(payload),
+        _effective_openai_max_tokens(payload),
         payload.stream,
         stop = payload.stop,
         min_p = payload.min_p,
@@ -11600,33 +12980,219 @@ async def _openai_passthrough_stream(
     completion_id,
     monitor_id: Optional[str] = None,
 ):
-    """Streaming client-side pass-through for /v1/chat/completions.
+    _cancel_keys = (payload.cancel_id, payload.session_id, completion_id)
+    _tracker = _TrackedCancel(cancel_event, *_cancel_keys)
+    _tracker.__enter__()
+    try:
+        reservation, admission_config = _openai_llama_admission_reserve(
+            request = request,
+            llama_backend = llama_backend,
+        )
+    except LlamaAdmissionQueueFull as exc:
+        _tracker.__exit__(None, None, None)
+        _openai_admission_log(
+            "queue-full",
+            snapshot = exc.snapshot,
+            request = request,
+            mode = "chat_passthrough_stream",
+            completion_id = completion_id,
+            level = "warning",
+        )
+        api_monitor.fail(monitor_id, str(exc))
+        raise _openai_admission_http_exception(exc, status_code = 429)
+
+    lease = reservation.lease_nowait()
+    if lease is not None:
+        try:
+            await _raise_if_openai_admission_cancelled(
+                reservation,
+                request = request,
+                cancel_event = cancel_event,
+            )
+        except asyncio.CancelledError:
+            api_monitor.finish(monitor_id, "cancelled")
+            lease.release()
+            _tracker.__exit__(None, None, None)
+            raise
+        except LlamaAdmissionCancelled as exc:
+            lease.release()
+            _tracker.__exit__(None, None, None)
+            api_monitor.finish(monitor_id, "cancelled")
+            raise HTTPException(
+                status_code = 499,
+                detail = _openai_admission_error_body(exc, status_code = 499),
+            )
+        return await _openai_passthrough_stream_admitted(
+            request,
+            cancel_event,
+            llama_backend,
+            payload,
+            model_name,
+            completion_id,
+            monitor_id = monitor_id,
+            admission_lease = lease,
+            tracker = _tracker,
+        )
+
+    admission_wait_started_at = time.monotonic()
+    _openai_admission_log(
+        "queued",
+        reservation,
+        request = request,
+        mode = "chat_passthrough_stream",
+        completion_id = completion_id,
+        level = "debug",
+    )
+
+    async def _queued_stream():
+        admitted_started = False
+        admitted_body_owns_cleanup = False
+        admitted_response = None
+        admitted_body_cancelled = False
+        try:
+            async for wait_item in _openai_admission_wait_stream_chunks(
+                reservation,
+                admission_config,
+                request = request,
+                cancel_event = cancel_event,
+            ):
+                if isinstance(wait_item, str):
+                    yield wait_item
+                    continue
+                _openai_admission_log(
+                    "granted-after-wait",
+                    reservation,
+                    request = request,
+                    mode = "chat_passthrough_stream",
+                    wait_started_at = admission_wait_started_at,
+                    completion_id = completion_id,
+                    level = "debug",
+                )
+                await _raise_if_openai_admission_cancelled(
+                    reservation,
+                    request = request,
+                    cancel_event = cancel_event,
+                )
+                admitted_response = await _openai_passthrough_stream_admitted(
+                    request,
+                    cancel_event,
+                    llama_backend,
+                    payload,
+                    model_name,
+                    completion_id,
+                    monitor_id = monitor_id,
+                    admission_lease = wait_item,
+                    tracker = _tracker,
+                )
+                admitted_started = True
+                iterator = admitted_response.body_iterator
+                admitted_body_owns_cleanup = True
+                try:
+                    async for chunk in iterator:
+                        yield chunk
+                except asyncio.CancelledError:
+                    admitted_body_cancelled = True
+                    raise
+                finally:
+                    await _close_openai_admitted_stream_iterator(
+                        iterator,
+                        cancelled = admitted_body_cancelled,
+                    )
+                    if not admitted_body_owns_cleanup:
+                        cleanup = getattr(admitted_response, "_unstarted_cleanup", None)
+                        if cleanup is not None:
+                            await cleanup()
+                return
+        except LlamaAdmissionTimeout as exc:
+            _openai_admission_log(
+                "timeout",
+                reservation,
+                request = request,
+                mode = "chat_passthrough_stream",
+                wait_started_at = admission_wait_started_at,
+                completion_id = completion_id,
+                level = "warning",
+            )
+            api_monitor.fail(monitor_id, str(exc))
+            yield _openai_stream_error_sse(_openai_admission_error_body(exc, status_code = 503))
+        except LlamaAdmissionCancelled:
+            _openai_admission_log(
+                "cancelled-before-upstream",
+                reservation,
+                request = request,
+                mode = "chat_passthrough_stream",
+                wait_started_at = admission_wait_started_at,
+                completion_id = completion_id,
+                level = "debug",
+            )
+            api_monitor.finish(monitor_id, "cancelled")
+            return
+        except asyncio.CancelledError:
+            api_monitor.finish(monitor_id, "cancelled")
+            raise
+        except HTTPException as exc:
+            status_code = getattr(exc, "status_code", 500) or 500
+            detail = exc.detail
+            error = (
+                detail
+                if isinstance(detail, dict) and "error" in detail
+                else openai_error_body(str(detail), status = status_code)
+            )
+            api_monitor.fail(monitor_id, str(detail))
+            yield _openai_stream_error_sse(error)
+        finally:
+            if not admitted_started:
+                api_monitor.finish(monitor_id, "cancelled")
+                reservation.cancel()
+                _tracker.__exit__(None, None, None)
+
+    async def _queued_unstarted_cleanup() -> None:
+        api_monitor.finish(monitor_id, "cancelled")
+        reservation.cancel()
+        _tracker.__exit__(None, None, None)
+
+    return _SameTaskStreamingResponse(
+        _queued_stream(),
+        media_type = "text/event-stream",
+        headers = {
+            "Cache-Control": "no-cache",
+            "Connection": "close",
+            "X-Accel-Buffering": "no",
+        },
+        unstarted_cleanup = _queued_unstarted_cleanup,
+    )
+
+
+async def _openai_passthrough_stream_admitted(
+    request,
+    cancel_event,
+    llama_backend,
+    payload,
+    model_name,
+    completion_id,
+    monitor_id: Optional[str] = None,
+    *,
+    admission_lease: LlamaAdmissionLease,
+    tracker,
+):
+    """Streaming client-side pass-through after Studio granted an upstream slot.
 
     Forwards the client's OpenAI function-calling request to llama-server and
-    relays the SSE stream back verbatim, preserving llama-server's native
-    response ``id``, ``finish_reason`` (including ``"tool_calls"``),
-    ``delta.tool_calls``, and any client-requested trailing ``usage`` chunk so
-    the client sees a standard OpenAI response.
+    relays the SSE stream back with minimal normalization (reasoning-only
+    deltas gain ``content: ""``; errors and missing terminal markers get a
+    closing ``[DONE]``), preserving llama-server's native response ``id``,
+    ``finish_reason`` (including ``"tool_calls"``), ``delta.tool_calls``, and
+    any client-requested trailing ``usage`` chunk so the client sees a
+    standard OpenAI response.
 
     Reasoning/tool-call splitting is delegated to llama-server (``--jinja
     --reasoning-format auto``), so ``delta.content`` carries no raw markup and is
     deliberately not re-parsed locally, unlike the ``/completion`` paths.
     """
+    _tracker = tracker
     target_url = f"{llama_backend.base_url}/v1/chat/completions"
-    body = _build_openai_passthrough_body(
-        payload, backend_ctx = llama_backend.context_length, llama_backend = llama_backend
-    )
-    # Text-form tool calls from small models get promoted to structured calls on
-    # the way back (declared client tools only); requests without tools or with
-    # auto_heal_tool_calls=false keep the verbatim relay. tool_choice constrains
-    # the allowlist ("none" disables, a forced function narrows to it).
-    _allowed_tools = heal_gate(
-        payload.auto_heal_tool_calls, body.get("tools"), body.get("tool_choice")
-    )
+    upstream_headers = _openai_passthrough_upstream_headers(llama_backend = llama_backend)
 
-    _cancel_keys = (payload.cancel_id, payload.session_id, completion_id)
-    _tracker = _TrackedCancel(cancel_event, *_cancel_keys)
-    _tracker.__enter__()
     client = None
     resp = None
     send_task: Optional[asyncio.Task[Optional[httpx.Response]]] = None
@@ -11648,6 +13214,17 @@ async def _openai_passthrough_stream(
 
     # Keep tracker cleanup paired if pre-header dispatch is cancelled.
     try:
+        body = _build_openai_passthrough_body(
+            payload, backend_ctx = llama_backend.context_length, llama_backend = llama_backend
+        )
+        # Text-form tool calls from small models get promoted to structured calls on
+        # the way back (declared client tools only); requests without tools or with
+        # auto_heal_tool_calls=false keep the unhealed relay. tool_choice constrains
+        # the allowlist ("none" disables, a forced function narrows to it).
+        _allowed_tools = heal_gate(
+            payload.auto_heal_tool_calls, body.get("tools"), body.get("tool_choice")
+        )
+
         # Keep the pre-header window short so accepted SSE clients receive
         # immediate headers in the common timeout-reduced stall.
         client = httpx.AsyncClient(
@@ -11661,12 +13238,16 @@ async def _openai_passthrough_stream(
 
         while True:
             try:
-                req = client.build_request(
-                    "POST", target_url, json = body, headers = {"Connection": "close"}
-                )
+                req = client.build_request("POST", target_url, json = body, headers = upstream_headers)
                 first_token_deadline = time.monotonic() + _DEFAULT_FIRST_TOKEN_TIMEOUT_S
                 send_task = asyncio.create_task(
-                    _send_stream_with_preheader_cancel(client, req, cancel_event, request = request)
+                    _send_stream_with_preheader_cancel(
+                        client,
+                        req,
+                        cancel_event,
+                        request = request,
+                        mark_cancel_on_cancel = False,
+                    )
                 )
                 done, _ = await asyncio.wait(
                     {send_task},
@@ -11692,13 +13273,17 @@ async def _openai_passthrough_stream(
             if resp is None and send_task is not None and not send_task.done():
                 break
             if resp is None:
+                if cancel_event is not None:
+                    cancel_event.set()
                 api_monitor.finish(monitor_id, "cancelled")
-                await _aclose_send_task(send_task)
                 try:
-                    await client.aclose()
-                except Exception:
-                    pass
-                _tracker.__exit__(None, None, None)
+                    await _aclose_send_task(send_task)
+                    await _aclose_stream_resources(client = client)
+                finally:
+                    try:
+                        admission_lease.release()
+                    finally:
+                        _tracker.__exit__(None, None, None)
                 return _SameTaskStreamingResponse(
                     iter(()),
                     media_type = "text/event-stream",
@@ -11723,6 +13308,7 @@ async def _openai_passthrough_stream(
                 await resp.aclose()
             except Exception:
                 pass
+            resp = None
             # Opt-in overflow policy: shrink and retry instead of a fatal 400.
             if (
                 _truncate_budget > 0
@@ -11751,11 +13337,14 @@ async def _openai_passthrough_stream(
             disconnect_watcher = None
 
             nonlocal resp, send_task, first_token_deadline, _truncate_budget
+            nonlocal client
             monitor_done = False
             saw_finish_reason = False
             saw_done = False
             saw_stream_error = False
+            saw_stream_item = False
             saw_tool_call_delta = False
+            terminal_seen = False
             last_chunk_id = completion_id
             last_chunk_model = model_name
             last_chunk_created = int(time.time())
@@ -11816,6 +13405,13 @@ async def _openai_passthrough_stream(
                     lines.append("data: " + json.dumps(chunk, ensure_ascii = False))
                 return lines
 
+            stall_timeout_s = _openai_compat_stream_stall_timeout()
+
+            def _terminal_read_timeout_s() -> Optional[float]:
+                if terminal_seen:
+                    return _OPENAI_PASSTHROUGH_TERMINAL_GRACE_S
+                return stall_timeout_s
+
             def _heal_transform(chunk_data: dict, raw_line: str) -> list:
                 """SSE lines to emit in place of one upstream line (healing on)."""
                 choices = chunk_data.get("choices")
@@ -11884,24 +13480,47 @@ async def _openai_passthrough_stream(
 
             try:
                 while True:
-                    if send_task is not None and not send_task.done():
-                        try:
-                            resp = await send_task
-                        except httpx.RequestError as e:
-                            logger.error("openai passthrough stream: upstream unreachable: %s", e)
-                            api_monitor.fail(monitor_id, _friendly_error(e))
-                            yield f"data: {json.dumps(_openai_stream_error_chunk(e))}\n\n"
-                            return
-                        send_task = None
-                    elif send_task is not None:
-                        try:
-                            resp = send_task.result()
-                        except httpx.RequestError as e:
-                            logger.error("openai passthrough stream: upstream unreachable: %s", e)
-                            api_monitor.fail(monitor_id, _friendly_error(e))
-                            yield f"data: {json.dumps(_openai_stream_error_chunk(e))}\n\n"
-                            return
-                        send_task = None
+                    if send_task is not None:
+                        last_keepalive_at = time.monotonic()
+                        while not send_task.done():
+                            # Wake often enough that _preheader_cancelled keeps
+                            # cancel/disconnect latency sub-second during prefill;
+                            # keepalives still pace off last_keepalive_at.
+                            wait_timeout = min(
+                                _STREAM_DISCONNECT_POLL_TIMEOUT_S,
+                                _OPENAI_PASSTHROUGH_PENDING_RESPONSE_KEEPALIVE_S,
+                            )
+                            done, _ = await asyncio.wait(
+                                {send_task},
+                                timeout = wait_timeout,
+                                return_when = asyncio.FIRST_COMPLETED,
+                            )
+                            if send_task in done:
+                                break
+                            if await _preheader_cancelled(cancel_event, request):
+                                api_monitor.finish(monitor_id, "cancelled")
+                                return
+                            # The downstream SSE response is already committed;
+                            # keep strict clients and proxies from treating a long
+                            # llama-server prefill/header wait as a dead stream.
+                            now = time.monotonic()
+                            if (
+                                now - last_keepalive_at
+                                >= _OPENAI_PASSTHROUGH_PENDING_RESPONSE_KEEPALIVE_S
+                            ):
+                                last_keepalive_at = now
+                                yield _OPENAI_PASSTHROUGH_SSE_KEEPALIVE
+                        if resp is None:
+                            try:
+                                resp = send_task.result()
+                            except httpx.RequestError as e:
+                                logger.error(
+                                    "openai passthrough stream: upstream unreachable: %s", e
+                                )
+                                api_monitor.fail(monitor_id, _friendly_error(e))
+                                yield _openai_stream_error_sse(_openai_stream_error_chunk(e))
+                                return
+                            send_task = None
 
                     if resp is None:
                         api_monitor.finish(monitor_id, "cancelled")
@@ -11929,12 +13548,16 @@ async def _openai_passthrough_stream(
                     ):
                         _truncate_budget -= 1
                         req = client.build_request(
-                            "POST", target_url, json = body, headers = {"Connection": "close"}
+                            "POST", target_url, json = body, headers = upstream_headers
                         )
                         first_token_deadline = time.monotonic() + _DEFAULT_FIRST_TOKEN_TIMEOUT_S
                         send_task = asyncio.create_task(
                             _send_stream_with_preheader_cancel(
-                                client, req, cancel_event, request = request
+                                client,
+                                req,
+                                cancel_event,
+                                request = request,
+                                mark_cancel_on_cancel = False,
                             )
                         )
                         continue
@@ -11949,7 +13572,7 @@ async def _openai_passthrough_stream(
                         )
                     )
                     api_monitor.fail(monitor_id, err_text[:500])
-                    yield f"data: {json.dumps(error_payload)}\n\n"
+                    yield _openai_stream_error_sse(error_payload)
                     return
 
                 cancel_watcher = asyncio.create_task(_await_cancel_then_close(cancel_event, resp))
@@ -11963,12 +13586,14 @@ async def _openai_passthrough_stream(
                     request = request,
                     first_token_deadline = first_token_deadline,
                     response = resp,
+                    post_first_item_read_timeout_s = _terminal_read_timeout_s,
                 ):
                     if not raw_line:
                         continue
-                    if not raw_line.startswith("data: "):
+                    if not raw_line.startswith("data:"):
                         continue
-                    data_text = raw_line[6:].strip()
+                    saw_stream_item = True
+                    data_text = raw_line[5:].strip()
                     if data_text == "[DONE]":
                         saw_done = True
                         # Upstream ended without a finish chunk: heal the residue
@@ -12000,13 +13625,11 @@ async def _openai_passthrough_stream(
                         yield raw_line + "\n\n"
                         monitor_done = True
                         break
-                    # Honor parallel_tool_calls=false (best-effort): drop tool_call
-                    # deltas with index>=1 so only the first call streams. Only
-                    # lines carrying tool_calls are reparsed; everything else is
-                    # relayed byte-for-byte.
-                    if payload.parallel_tool_calls is False and '"tool_calls"' in raw_line:
-                        raw_line = _cap_parallel_tool_calls_sse_line(raw_line)
-                        data_text = raw_line[6:].strip()
+                    raw_line = _normalize_openai_passthrough_sse_line(
+                        raw_line,
+                        cap_parallel_tool_calls = payload.parallel_tool_calls is False,
+                    )
+                    data_text = raw_line[5:].strip()
                     try:
                         chunk_data = json.loads(data_text)
                     except json.JSONDecodeError:
@@ -12033,8 +13656,9 @@ async def _openai_passthrough_stream(
                         if _monitor_openai_error_message(chunk_data):
                             saw_stream_error = True
                     # With healing active, a content-bearing line may be replaced by
-                    # held/promoted chunks; otherwise the single upstream line
-                    # relays verbatim (monitored exactly as emitted either way).
+                    # held/promoted chunks; otherwise the single (already
+                    # normalized) line relays unchanged (monitored exactly as
+                    # emitted either way).
                     if (
                         healer is not None
                         and not healer.dormant
@@ -12084,6 +13708,27 @@ async def _openai_passthrough_stream(
                         yield out_line + "\n\n"
                         if monitor_event == "done":
                             monitor_done = True
+                            break
+                        terminal_state = (
+                            _openai_passthrough_terminal_state_from_data(chunk_data)
+                            if out_line is raw_line
+                            else _openai_passthrough_sse_line_terminal_state(out_line)
+                        )
+                        if terminal_state == "usage" or (
+                            terminal_state == "finish" and not _wants_stream_usage(payload)
+                        ):
+                            done_line = _SSE_DONE_LINE
+                            _monitor_openai_sse_line(
+                                monitor_id,
+                                done_line,
+                                llama_backend.context_length,
+                            )
+                            yield done_line + "\n\n"
+                            saw_done = True
+                            monitor_done = True
+                            break
+                        if terminal_state == "finish":
+                            terminal_seen = True
                     if monitor_done:
                         break
                 if not saw_done and not saw_stream_error and not cancel_event.is_set():
@@ -12105,7 +13750,7 @@ async def _openai_passthrough_stream(
                             llama_backend.context_length,
                         )
                         yield finish_line + "\n\n"
-                    done_line = "data: [DONE]"
+                    done_line = _SSE_DONE_LINE
                     _monitor_openai_sse_line(
                         monitor_id,
                         done_line,
@@ -12121,6 +13766,29 @@ async def _openai_passthrough_stream(
             except asyncio.CancelledError:
                 api_monitor.finish(monitor_id, "cancelled")
                 raise
+            except httpx.ReadTimeout as e:
+                if terminal_seen and not saw_stream_error and not cancel_event.is_set():
+                    done_line = _SSE_DONE_LINE
+                    _monitor_openai_sse_line(
+                        monitor_id,
+                        done_line,
+                        llama_backend.context_length,
+                    )
+                    yield done_line + "\n\n"
+                    api_monitor.finish(monitor_id)
+                    return
+                if cancel_event.is_set():
+                    api_monitor.finish(monitor_id, "cancelled")
+                    return
+                logger.error(
+                    "openai passthrough stream %s: %s",
+                    "stalled mid-response" if saw_stream_item else "timeout",
+                    e,
+                )
+                api_monitor.fail(monitor_id, _friendly_error(e))
+                get_llama_cpp_backend()._maybe_recover_from_mtp_crash(e)
+                err = _openai_stream_error_chunk(e)
+                yield _openai_stream_error_sse(err)
             except (httpx.RemoteProtocolError, httpx.ReadError, httpx.CloseError) as e:
                 # Watcher closed resp on cancel. Emit nothing extra; the client
                 # initiated the cancel or already disconnected.
@@ -12129,6 +13797,16 @@ async def _openai_passthrough_stream(
                     get_llama_cpp_backend()._maybe_recover_from_mtp_crash(e)
                     raise
                 api_monitor.finish(monitor_id, "cancelled")
+            except HTTPException as exc:
+                status_code = getattr(exc, "status_code", 500) or 500
+                detail = exc.detail
+                error_payload = (
+                    detail
+                    if isinstance(detail, dict) and "error" in detail
+                    else openai_error_body(str(detail), status = status_code)
+                )
+                api_monitor.fail(monitor_id, str(detail))
+                yield _openai_stream_error_sse(error_payload)
             except Exception as e:
                 if cancel_event.is_set():
                     api_monitor.finish(monitor_id, "cancelled")
@@ -12138,25 +13816,37 @@ async def _openai_passthrough_stream(
                 api_monitor.fail(monitor_id, _friendly_error(e))
                 get_llama_cpp_backend()._maybe_recover_from_mtp_crash(e)
                 err = _openai_stream_error_chunk(e)
-                yield f"data: {json.dumps(err)}\n\n"
+                yield _openai_stream_error_sse(err)
             finally:
-                await _aclose_send_task(send_task)
-                await _aclose_stream_resources(
-                    watchers = (cancel_watcher, disconnect_watcher),
-                    iterator = lines_iter,
-                    resp = resp,
-                    client = client,
-                )
-                _tracker.__exit__(None, None, None)
+                # _aclose_stream_resources re-raises a close-time CancelledError
+                # only after finishing teardown, and the tracker exits either way.
+                try:
+                    await _aclose_send_task(send_task)
+                    await _aclose_stream_resources(
+                        watchers = (cancel_watcher, disconnect_watcher),
+                        iterator = lines_iter,
+                        resp = resp,
+                        client = client,
+                    )
+                finally:
+                    try:
+                        admission_lease.release()
+                    finally:
+                        _tracker.__exit__(None, None, None)
 
         async def _unstarted_cleanup() -> None:
             # Client disconnected before the body stream started, so _stream()'s
             # finally never ran. Release the eagerly-opened upstream resp/client
             # and the cancel-registry entry here; the watchers and line iterator
             # are created inside _stream(), so there is nothing else to close.
-            await _aclose_send_task(send_task)
-            await _aclose_stream_resources(resp = resp, client = client)
-            _tracker.__exit__(None, None, None)
+            try:
+                await _aclose_send_task(send_task)
+                await _aclose_stream_resources(resp = resp, client = client)
+            finally:
+                try:
+                    admission_lease.release()
+                finally:
+                    _tracker.__exit__(None, None, None)
 
         return _SameTaskStreamingResponse(
             _stream(),
@@ -12168,10 +13858,22 @@ async def _openai_passthrough_stream(
             },
             unstarted_cleanup = _unstarted_cleanup,
         )
-    except BaseException:
-        await _aclose_send_task(send_task)
-        await _aclose_stream_resources(resp = resp, client = client)
-        _tracker.__exit__(None, None, None)
+    except BaseException as exc:
+        if isinstance(exc, asyncio.CancelledError):
+            if cancel_event is not None:
+                cancel_event.set()
+            api_monitor.finish(monitor_id, "cancelled")
+        else:
+            detail = exc.detail if isinstance(exc, HTTPException) else _friendly_error(exc)
+            api_monitor.fail(monitor_id, str(detail))
+        try:
+            await _aclose_send_task(send_task)
+            await _aclose_stream_resources(resp = resp, client = client)
+        finally:
+            try:
+                admission_lease.release()
+            finally:
+                _tracker.__exit__(None, None, None)
         raise
 
 
@@ -12180,6 +13882,109 @@ async def _openai_passthrough_non_streaming(
     payload,
     model_name,
     monitor_id: Optional[str] = None,
+    *,
+    request: Optional[Request] = None,
+    cancel_event = None,
+):
+    """Non-streaming pass-through guarded by local llama-server admission."""
+    try:
+        reservation, admission_config = _openai_llama_admission_reserve(
+            request = request,
+            llama_backend = llama_backend,
+        )
+    except LlamaAdmissionQueueFull as exc:
+        _openai_admission_log(
+            "queue-full",
+            snapshot = exc.snapshot,
+            request = request,
+            mode = "chat_passthrough_nonstream",
+            level = "warning",
+        )
+        api_monitor.fail(monitor_id, str(exc))
+        raise _openai_admission_http_exception(exc, status_code = 429)
+
+    lease = None
+    admission_wait_started_at = None
+    try:
+        if reservation.lease_nowait() is None:
+            admission_wait_started_at = time.monotonic()
+            _openai_admission_log(
+                "queued",
+                reservation,
+                request = request,
+                mode = "chat_passthrough_nonstream",
+                level = "debug",
+            )
+        lease = await _wait_for_openai_admission_non_streaming(
+            reservation,
+            admission_config,
+            request = request,
+            cancel_event = cancel_event,
+        )
+        if admission_wait_started_at is not None:
+            _openai_admission_log(
+                "granted-after-wait",
+                reservation,
+                request = request,
+                mode = "chat_passthrough_nonstream",
+                wait_started_at = admission_wait_started_at,
+                level = "debug",
+            )
+        await _raise_if_openai_admission_cancelled(
+            reservation,
+            request = request,
+            cancel_event = cancel_event,
+        )
+        return await _openai_passthrough_non_streaming_upstream(
+            llama_backend,
+            payload,
+            model_name,
+            monitor_id = monitor_id,
+            request = request,
+            cancel_event = cancel_event,
+        )
+    except LlamaAdmissionTimeout as exc:
+        _openai_admission_log(
+            "timeout",
+            reservation,
+            request = request,
+            mode = "chat_passthrough_nonstream",
+            wait_started_at = admission_wait_started_at,
+            level = "warning",
+        )
+        api_monitor.fail(monitor_id, str(exc))
+        raise _openai_admission_http_exception(exc, status_code = 503)
+    except LlamaAdmissionCancelled as exc:
+        _openai_admission_log(
+            "cancelled-before-upstream",
+            reservation,
+            request = request,
+            mode = "chat_passthrough_nonstream",
+            wait_started_at = admission_wait_started_at,
+            level = "debug",
+        )
+        api_monitor.finish(monitor_id, "cancelled")
+        raise HTTPException(
+            status_code = 499,
+            detail = _openai_admission_error_body(exc, status_code = 499),
+        )
+    except asyncio.CancelledError:
+        api_monitor.finish(monitor_id, "cancelled")
+        reservation.cancel()
+        raise
+    finally:
+        if lease is not None:
+            lease.release()
+
+
+async def _openai_passthrough_non_streaming_upstream(
+    llama_backend,
+    payload,
+    model_name,
+    monitor_id: Optional[str] = None,
+    *,
+    request: Optional[Request] = None,
+    cancel_event = None,
 ):
     """Non-streaming client-side pass-through for /v1/chat/completions.
 
@@ -12188,20 +13993,67 @@ async def _openai_passthrough_non_streaming(
     ``tool_calls``, and accurate ``usage`` token counts.
     """
     target_url = f"{llama_backend.base_url}/v1/chat/completions"
+    upstream_headers = _openai_passthrough_upstream_headers(llama_backend = llama_backend)
     body = _build_openai_passthrough_body(
         payload, backend_ctx = llama_backend.context_length, llama_backend = llama_backend
     )
+    body["stream"] = False
+    body.pop("stream_options", None)
 
     _truncate_budget = (
         _OVERFLOW_TRUNCATE_MAX_RETRIES if _overflow_truncation_requested(payload) else 0
     )
-    while True:
-        try:
-            resp = await nonstreaming_client().post(
+
+    async def _post(body_to_send):
+        if cancel_event is None and request is None:
+            return await nonstreaming_client().post(
                 target_url,
-                json = body,
+                json = body_to_send,
+                headers = upstream_headers,
                 timeout = _llama_non_streaming_generation_timeout(),
             )
+
+        if cancel_event is None:
+            cancel = threading.Event()
+        else:
+            cancel = cancel_event
+        client = _cancelable_nonstreaming_client()
+        watcher = asyncio.create_task(
+            _await_cancel_or_disconnect_then_close_client(
+                cancel_event = cancel,
+                request = request,
+                client = client,
+            )
+        )
+        try:
+            try:
+                response = await client.post(
+                    target_url,
+                    json = body_to_send,
+                    headers = upstream_headers,
+                    timeout = _llama_non_streaming_generation_timeout(),
+                )
+            except httpx.RequestError:
+                if cancel.is_set():
+                    raise asyncio.CancelledError()
+                raise
+            if cancel.is_set():
+                raise asyncio.CancelledError()
+            return response
+        finally:
+            watcher.cancel()
+            try:
+                await watcher
+            except (asyncio.CancelledError, Exception):
+                pass
+            try:
+                await client.aclose()
+            except Exception:
+                pass
+
+    while True:
+        try:
+            resp = await _post(body)
         except asyncio.CancelledError:
             api_monitor.finish(monitor_id, "cancelled")
             raise
@@ -12269,15 +14121,14 @@ async def _openai_passthrough_non_streaming(
             "messages": [*body.get("messages", []), *nudge_messages(data, _allowed_tools)],
         }
         try:
-            retry_resp = await nonstreaming_client().post(
-                target_url,
-                json = retry_body,
-                timeout = _llama_non_streaming_generation_timeout(),
-            )
+            retry_resp = await _post(retry_body)
             if retry_resp.status_code == 200:
                 retry_data = retry_resp.json()
                 if response_has_promotable_calls(retry_data, _allowed_tools, body.get("tools")):
                     resp, data = retry_resp, retry_data
+        except asyncio.CancelledError:
+            api_monitor.finish(monitor_id, "cancelled")
+            raise
         except (httpx.RequestError, ValueError) as exc:
             logger.warning("tool-call nudge retry failed; keeping original: %s", exc)
 
diff --git a/studio/backend/routes/settings.py b/studio/backend/routes/settings.py
index 914699f540..bbee374334 100644
--- a/studio/backend/routes/settings.py
+++ b/studio/backend/routes/settings.py
@@ -32,6 +32,7 @@ from utils.helper_precache_settings import (
     helper_model_disabled_by_env,
     set_helper_precache_enabled,
 )
+from utils.coding_agents import CODING_AGENTS, detect_installed_coding_agents
 from utils.openai_auto_switch_settings import (
     DEFAULT_AUTO_UNLOAD_IDLE_SECONDS,
     DEFAULT_OPENAI_AUTO_SWITCH_ENABLED,
@@ -174,6 +175,19 @@ def update_helper_precache(
     return _helper_precache_response(enabled)
 
 
+class CodingAgentsResponse(BaseModel):
+    # All agents `unsloth start` supports, in the CLI's declared order.
+    agents: tuple[str, ...] = CODING_AGENTS
+    # Subset of `agents` whose CLI binary was found on PATH; the frontend uses
+    # this to default the API-keys panel to a command the user can run as-is.
+    detected: list[str]
+
+
+@router.get("/coding-agents", response_model = CodingAgentsResponse)
+def get_coding_agents(current_subject: str = Depends(get_current_subject)) -> CodingAgentsResponse:
+    return CodingAgentsResponse(detected = detect_installed_coding_agents())
+
+
 @router.get("/openai-auto-switch", response_model = OpenAIAutoSwitchResponse)
 def get_openai_auto_switch(
     current_subject: str = Depends(get_current_subject),
diff --git a/studio/backend/tests/test_anthropic_messages.py b/studio/backend/tests/test_anthropic_messages.py
index 0c6550a3bb..54454f6563 100644
--- a/studio/backend/tests/test_anthropic_messages.py
+++ b/studio/backend/tests/test_anthropic_messages.py
@@ -1770,3 +1770,187 @@ class TestAnthropicMessagesToolRouting:
 
         _drive(anthropic_messages(payload, request = None, current_subject = "t"))
         assert backend.calls[0][0] == "plain"
+
+
+def test_resumed_session_thinking_and_null_content_do_not_400():
+    # A resumed session replays assistant turns with `thinking` (and sometimes null)
+    # content. Those must be accepted (thinking dropped by the converter), not 400ed.
+    from pydantic import ValidationError
+
+    req = AnthropicMessagesRequest(
+        model = "x",
+        max_tokens = 16,
+        messages = [
+            {"role": "user", "content": "hi"},
+            {
+                "role": "assistant",
+                "content": [
+                    {"type": "thinking", "thinking": "secret reasoning", "signature": "s"},
+                    {"type": "text", "text": "the answer"},
+                    {"type": "tool_use", "id": "t1", "name": "f", "input": {}},
+                ],
+            },
+            {"role": "assistant", "content": None},  # tool-only turn serialized as null
+        ],
+    )
+    # Known blocks still parse as their typed models; only the unknown one is loose.
+    assert type(req.messages[1].content[0]).__name__ == "AnthropicUnknownBlock"
+    assert type(req.messages[1].content[1]).__name__ == "AnthropicTextBlock"
+    assert req.messages[2].content == ""  # null coerced
+
+    openai = anthropic_messages_to_openai([m.model_dump() for m in req.messages])
+    assistant = next(m for m in openai if m["role"] == "assistant" and m.get("content"))
+    assert assistant["content"] == "the answer"
+    assert "secret reasoning" not in json.dumps(openai)  # thinking never forwarded
+
+    # A malformed KNOWN block still fails cleanly instead of being swallowed.
+    with pytest.raises(ValidationError):
+        AnthropicMessagesRequest(
+            model = "x",
+            max_tokens = 16,
+            messages = [{"role": "assistant", "content": [{"type": "tool_use", "name": "f"}]}],
+        )
+
+
+def test_user_null_content_rejected():
+    # The null->"" leniency is assistant-only; a null user content must be rejected
+    # at the boundary, not coerced into an empty prompt and forwarded to the model.
+    from pydantic import ValidationError
+    with pytest.raises(ValidationError):
+        AnthropicMessagesRequest(
+            model = "x",
+            max_tokens = 16,
+            messages = [{"role": "user", "content": None}],
+        )
+
+
+def test_user_unknown_block_rejected_not_silently_dropped():
+    # The converter skips user blocks it cannot translate, so a user turn whose only
+    # block is unknown would validate yet forward no content. Reject at the boundary
+    # to avoid that silent data loss (the assistant fallback is unaffected).
+    from pydantic import ValidationError
+    with pytest.raises(ValidationError):
+        AnthropicMessagesRequest(
+            model = "x",
+            max_tokens = 16,
+            messages = [
+                {"role": "user", "content": [{"type": "document", "source": {}}]},
+            ],
+        )
+
+
+def test_user_translatable_blocks_still_accepted():
+    # text / image / tool_result are translatable, so a real user message built from
+    # them must still pass; the unknown-block guard only trips on other types.
+    req = AnthropicMessagesRequest(
+        model = "x",
+        max_tokens = 16,
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "What is this?"},
+                    {
+                        "type": "image",
+                        "source": {"type": "base64", "media_type": "image/png", "data": "AA"},
+                    },
+                    {"type": "tool_result", "tool_use_id": "t1", "content": "ok"},
+                ],
+            }
+        ],
+    )
+    assert [type(b).__name__ for b in req.messages[0].content] == [
+        "AnthropicTextBlock",
+        "AnthropicImageBlock",
+        "AnthropicToolResultBlock",
+    ]
+
+    openai = anthropic_messages_to_openai([m.model_dump() for m in req.messages])
+    assert any(m["role"] == "tool" and m["tool_call_id"] == "t1" for m in openai)
+
+
+def test_user_malformed_known_block_still_rejected():
+    # The guard only allow-lists a user block's *type*; the union still validates its
+    # shape, so a known-but-malformed block (tool_result without tool_use_id) fails.
+    from pydantic import ValidationError
+    with pytest.raises(ValidationError):
+        AnthropicMessagesRequest(
+            model = "x",
+            max_tokens = 16,
+            messages = [
+                {"role": "user", "content": [{"type": "tool_result", "content": "x"}]},
+            ],
+        )
+
+
+def test_user_content_block_non_string_type_rejected_cleanly():
+    # A user block whose `type` is a non-string (unhashable list / dict, or a stray
+    # int) must fail as a clean validation error, not raise TypeError from the
+    # frozenset membership test and escape as a 500.
+    from pydantic import ValidationError
+    for bad_type in ([], {}, 5):
+        with pytest.raises(ValidationError):
+            AnthropicMessagesRequest(
+                model = "x",
+                max_tokens = 16,
+                messages = [{"role": "user", "content": [{"type": bad_type}]}],
+            )
+
+
+def test_assistant_missing_content_key_still_rejected():
+    # The null -> "" leniency is only for an EXPLICIT null. An assistant message that
+    # omits content entirely stays malformed and must fail required-field validation.
+    from pydantic import ValidationError
+
+    with pytest.raises(ValidationError):
+        AnthropicMessagesRequest(
+            model = "x",
+            max_tokens = 16,
+            messages = [{"role": "assistant"}],
+        )
+    # An explicit null is still accepted and coerced (regression guard).
+    req = AnthropicMessagesRequest(
+        model = "x",
+        max_tokens = 16,
+        messages = [
+            {"role": "user", "content": "hi"},
+            {"role": "assistant", "content": None},
+        ],
+    )
+    assert req.messages[1].content == ""
+
+
+def test_resumed_null_assistant_between_users_coalesced_on_messages_route(monkeypatch):
+    # user -> assistant(null) -> user is now accepted: the null assistant turn coerces
+    # to "" and is dropped. The route must then coalesce the two remaining user turns
+    # so a strict GGUF chat template does not 400 on non-alternating roles.
+    backend = _mock_backend(monkeypatch, context_length = 2048)
+
+    class _Req:
+        state = SimpleNamespace()
+        url = SimpleNamespace(path = "/v1/messages")
+        method = "POST"
+
+        async def is_disconnected(self):
+            return False
+
+    payload = AnthropicMessagesRequest(
+        model = "x",
+        max_tokens = 16,
+        messages = [
+            {"role": "user", "content": "first question"},
+            {"role": "assistant", "content": None},
+            {"role": "user", "content": "please continue"},
+        ],
+    )
+
+    response = _drive(anthropic_messages(payload, request = _Req(), current_subject = "t"))
+    assert response.status_code == 200
+
+    [(_path, kwargs)] = backend.calls
+    user_turns = [m for m in kwargs["messages"] if m.get("role") == "user"]
+    assert len(user_turns) == 1  # the two user turns were merged, not left adjacent
+    merged = user_turns[0]["content"]
+    if isinstance(merged, list):
+        merged = " ".join(p.get("text", "") for p in merged if isinstance(p, dict))
+    assert "first question" in merged and "please continue" in merged
diff --git a/studio/backend/tests/test_coding_agents.py b/studio/backend/tests/test_coding_agents.py
new file mode 100644
index 0000000000..b19da1dded
--- /dev/null
+++ b/studio/backend/tests/test_coding_agents.py
@@ -0,0 +1,50 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Tests for coding-agent CLI detection used by the API-keys settings panel."""
+
+from unittest.mock import patch
+
+from utils.coding_agents import CODING_AGENTS, detect_installed_coding_agents
+
+
+def test_matches_unsloth_start_subcommands():
+    # Each entry must be an actual `unsloth start <agent>` subcommand name
+    # (unsloth_cli/commands/start.py). Spelled out here rather than imported
+    # from that module, which pulls in the CLI's heavier dependencies.
+    assert CODING_AGENTS == ("claude", "codex", "openclaw", "opencode", "hermes", "pi")
+
+
+def test_detects_only_agents_present_on_path():
+    installed = {"claude", "opencode"}
+    with patch(
+        "utils.coding_agents.shutil.which",
+        side_effect = lambda name: f"/usr/bin/{name}" if name in installed else None,
+    ):
+        assert detect_installed_coding_agents() == ["claude", "opencode"]
+
+
+def test_returns_empty_list_when_nothing_is_installed():
+    with patch("utils.coding_agents.shutil.which", return_value = None):
+        assert detect_installed_coding_agents() == []
+
+
+def test_preserves_declared_order_regardless_of_path_lookup_order():
+    with patch(
+        "utils.coding_agents.shutil.which",
+        side_effect = lambda name: name if name in ("pi", "claude", "hermes") else None,
+    ):
+        assert detect_installed_coding_agents() == ["claude", "hermes", "pi"]
+
+
+def test_treats_a_path_lookup_error_as_not_installed():
+    # An advisory check: shutil.which raising for one entry (e.g. a permission
+    # error walking a PATH directory) should not take down the whole endpoint,
+    # and should not stop the remaining agents from being checked.
+    def flaky_which(name: str):
+        if name == "codex":
+            raise OSError("permission denied")
+        return name if name == "claude" else None
+
+    with patch("utils.coding_agents.shutil.which", side_effect = flaky_which):
+        assert detect_installed_coding_agents() == ["claude"]
diff --git a/studio/backend/tests/test_completion_masking.py b/studio/backend/tests/test_completion_masking.py
new file mode 100644
index 0000000000..be0d8a69bd
--- /dev/null
+++ b/studio/backend/tests/test_completion_masking.py
@@ -0,0 +1,314 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Completion-only masking policy: auto-detect first, manual table fallback.
+
+Covers utils.datasets.completion_masking.apply_completion_masking, shared by
+the CUDA trainer (core/training/trainer.py) and the MLX worker
+(core/training/worker.py):
+  - unmapped models use chat template auto-detection (previously masking was
+    silently disabled),
+  - gpt-oss goes auto-first too (its quantized checkpoints ship a template
+    the manual markers cannot match),
+  - an auto-detection failure falls back to the template table markers,
+  - a table miss after an auto failure warns and leaves the trainer unchanged.
+"""
+
+from __future__ import annotations
+
+import pytest
+
+from utils.datasets.completion_masking import apply_completion_masking, lookup_manual_markers
+from utils.datasets.model_mappings import TEMPLATE_TO_RESPONSES_MAPPER
+
+
+class _Trainer:
+    """Sentinel trainer; train_fn wraps it in a new object when applied."""
+
+
+class _Recorder:
+    """Fake train_on_responses_only that records calls."""
+
+    def __init__(self):
+        self.calls = []
+
+    def __call__(self, trainer, **kwargs):
+        self.calls.append(kwargs)
+        wrapped = _Trainer()
+        wrapped.wrapped_from = trainer
+        return wrapped
+
+
+def _detect_ok(processor):
+    return "<INS>", "<RES>"
+
+
+def _detect_fail(processor):
+    raise ValueError(
+        "Unsloth: Could not reliably auto-detect response_part - "
+        "pass instruction_part and response_part."
+    )
+
+
+_AUTO = {"instruction_part": "<INS>", "response_part": "<RES>"}
+
+
+class _Notes:
+    def __init__(self):
+        self.messages = []
+
+    def __call__(self, level, message):
+        self.messages.append((level, message))
+
+    def warnings(self):
+        return [m for level, m in self.messages if level == "warning"]
+
+
+def test_unmapped_model_uses_auto_detection():
+    # Unmapped model: the auto path applies masking (was silently disabled).
+    trainer = _Trainer()
+    train_fn = _Recorder()
+    notes = _Notes()
+
+    result, applied = apply_completion_masking(
+        trainer, "LiquidAI/LFM2-8B-A1B", train_fn, notify = notes, detect_fn = _detect_ok
+    )
+
+    assert applied is True
+    assert result.wrapped_from is trainer
+    assert train_fn.calls == [dict(_AUTO)]  # applied with the detected markers
+    assert notes.warnings() == []
+
+
+def test_mapped_model_prefers_auto_detection():
+    trainer = _Trainer()
+    train_fn = _Recorder()
+
+    _, applied = apply_completion_masking(
+        trainer, "unsloth/Qwen3-0.6B", train_fn, detect_fn = _detect_ok
+    )
+
+    assert applied is True
+    assert train_fn.calls == [dict(_AUTO)]
+
+
+def test_gpt_oss_uses_auto_detection_first():
+    # The quantized gpt-oss checkpoints ship a template without the
+    # <|channel|>final header, where the manual markers match nothing; auto
+    # derives markers from the template the checkpoint actually ships.
+    trainer = _Trainer()
+    train_fn = _Recorder()
+
+    _, applied = apply_completion_masking(
+        trainer, "unsloth/gpt-oss-20b", train_fn, detect_fn = _detect_ok
+    )
+
+    assert applied is True
+    assert train_fn.calls == [dict(_AUTO)]
+
+
+def test_gpt_oss_detection_failure_falls_back_to_manual_markers():
+    trainer = _Trainer()
+    train_fn = _Recorder()
+
+    _, applied = apply_completion_masking(
+        trainer, "unsloth/gpt-oss-20b", train_fn, detect_fn = _detect_fail
+    )
+
+    assert applied is True
+    expected = TEMPLATE_TO_RESPONSES_MAPPER["gpt-oss"]
+    assert train_fn.calls == [
+        {
+            "instruction_part": expected["instruction"],
+            "response_part": expected["response"],
+        }
+    ]
+
+
+def test_auto_failure_falls_back_to_template_table():
+    trainer = _Trainer()
+    train_fn = _Recorder()
+    notes = _Notes()
+
+    result, applied = apply_completion_masking(
+        trainer, "unsloth/Qwen3-0.6B", train_fn, notify = notes, detect_fn = _detect_fail
+    )
+
+    assert applied is True
+    assert result.wrapped_from is trainer
+    expected = TEMPLATE_TO_RESPONSES_MAPPER["qwen3"]
+    assert train_fn.calls == [
+        {
+            "instruction_part": expected["instruction"],
+            "response_part": expected["response"],
+        },
+    ]
+    assert any("falling back to the template table" in m for m in notes.warnings())
+
+
+def test_application_failure_propagates_not_fallback():
+    # Detection succeeds; a failure while APPLYING the masking must propagate,
+    # never silently fall back to full-sequence training.
+    def train_fn(trainer, **kwargs):
+        raise RuntimeError("dataset map worker crashed")
+
+    with pytest.raises(RuntimeError, match = "dataset map worker crashed"):
+        apply_completion_masking(_Trainer(), "LiquidAI/LFM2-8B-A1B", train_fn, detect_fn = _detect_ok)
+
+
+def test_preset_tokenizer_markers_used_directly():
+    # Preset unsloth marker attrs skip detection; zoo reuses them on a bare call.
+    class _Tok:
+        _unsloth_input_part = "<I>"
+        _unsloth_output_part = "<O>"
+
+    trainer = _Trainer()
+    trainer.processing_class = _Tok()
+    train_fn = _Recorder()
+
+    _, applied = apply_completion_masking(
+        trainer, "LiquidAI/LFM2-8B-A1B", train_fn, detect_fn = _detect_fail
+    )
+    assert applied is True
+    assert train_fn.calls == [{}]  # bare call, stored parts
+
+
+def test_table_miss_warns_and_disables_without_crashing():
+    trainer = _Trainer()
+    train_fn = _Recorder()
+    notes = _Notes()
+
+    result, applied = apply_completion_masking(
+        trainer, "some-org/not-in-any-mapper", train_fn, notify = notes, detect_fn = _detect_fail
+    )
+
+    assert applied is False
+    assert result is trainer  # unchanged: full sequence training
+    assert train_fn.calls == []  # detection failed; nothing applied
+    assert any("could not be applied" in m for m in notes.warnings())
+    assert any("full sequences" in m for m in notes.warnings())
+
+
+def test_num_proc_forwarded_only_when_given():
+    # CUDA path passes num_proc; the MLX path omits it.
+    train_fn = _Recorder()
+    apply_completion_masking(
+        _Trainer(), "unsloth/Qwen3-0.6B", train_fn, num_proc = 4, detect_fn = _detect_ok
+    )
+    assert train_fn.calls == [dict(_AUTO, num_proc = 4)]
+
+    train_fn = _Recorder()
+    apply_completion_masking(
+        _Trainer(), "unsloth/Qwen3-0.6B", train_fn, num_proc = 4, detect_fn = _detect_fail
+    )
+    assert train_fn.calls[0]["num_proc"] == 4
+
+    train_fn = _Recorder()
+    apply_completion_masking(_Trainer(), "unsloth/Qwen3-0.6B", train_fn, detect_fn = _detect_ok)
+    assert train_fn.calls == [dict(_AUTO)]
+
+
+def test_manual_fallback_failure_propagates_to_caller():
+    # Errors while applying the manual fallback must propagate to the caller.
+    def train_fn(trainer, **kwargs):
+        raise RuntimeError("boom")
+
+    with pytest.raises(RuntimeError, match = "boom"):
+        apply_completion_masking(_Trainer(), "unsloth/gpt-oss-20b", train_fn)
+
+
+def test_notify_is_optional():
+    train_fn = _Recorder()
+    _, applied = apply_completion_masking(
+        _Trainer(), "some-org/not-in-any-mapper", train_fn, detect_fn = _detect_fail
+    )
+    assert applied is False
+
+
+def test_lookup_manual_markers():
+    template, instruction, response = lookup_manual_markers("unsloth/Qwen3-0.6B")
+    assert template == "qwen3"
+    assert instruction == TEMPLATE_TO_RESPONSES_MAPPER["qwen3"]["instruction"]
+    assert response == TEMPLATE_TO_RESPONSES_MAPPER["qwen3"]["response"]
+
+    template, instruction, response = lookup_manual_markers("some-org/unknown")
+    assert (template, instruction, response) == (None, None, None)
+
+    template, instruction, response = lookup_manual_markers(None)
+    assert (template, instruction, response) == (None, None, None)
+
+
+def test_renamed_gpt_oss_gets_template_markers():
+    # Name-detected as gpt-oss but not in the exact-name table: must use the
+    # gpt-oss markers, not fall through to full-sequence training.
+    trainer = _Trainer()
+    train_fn = _Recorder()
+
+    _, applied = apply_completion_masking(
+        trainer, "some-org/gpt-oss-20b-sft", train_fn, detect_fn = _detect_fail
+    )
+    assert applied is True
+    expected = TEMPLATE_TO_RESPONSES_MAPPER["gpt-oss"]
+    assert train_fn.calls == [
+        {
+            "instruction_part": expected["instruction"],
+            "response_part": expected["response"],
+        }
+    ]
+
+
+class _FakeTokenizerWrapper:
+    """mlx-lm TokenizerWrapper semantics: plain reads delegate to the wrapped
+    tokenizer, underscore attrs do not (so preset markers are hidden)."""
+
+    def __init__(self, tokenizer):
+        object.__setattr__(self, "_tokenizer", tokenizer)
+
+    def __getattr__(self, attr):
+        if attr.startswith("_"):
+            return object.__getattribute__(self, attr)
+        return getattr(object.__getattribute__(self, "_tokenizer"), attr)
+
+
+_FakeTokenizerWrapper.__name__ = "TokenizerWrapper"
+
+
+def test_mlx_tokenizer_wrapper_unwrapped_for_preset_markers():
+    # Markers live on the inner HF tokenizer that the wrapper hides; the helper
+    # must unwrap so the preset bare-call path still fires on MLX.
+    class _Tok:
+        _unsloth_input_part = "<I>"
+        _unsloth_output_part = "<O>"
+
+    trainer = _Trainer()
+    trainer.tokenizer = _FakeTokenizerWrapper(_Tok())
+    train_fn = _Recorder()
+
+    _, applied = apply_completion_masking(
+        trainer, "LiquidAI/LFM2-8B-A1B", train_fn, detect_fn = _detect_fail
+    )
+    assert applied is True
+    assert train_fn.calls == [{}]  # bare call, stored parts
+
+
+def test_mlx_tokenizer_wrapper_unwrapped_for_detection():
+    # Detection must see the real tokenizer, not the wrapper, so it does not
+    # depend on the loader's __call__ patch.
+    class _Tok:
+        pass
+
+    inner = _Tok()
+    trainer = _Trainer()
+    trainer.tokenizer = _FakeTokenizerWrapper(inner)
+    train_fn = _Recorder()
+    seen = []
+
+    def detect(processor):
+        seen.append(processor)
+        return "<INS>", "<RES>"
+
+    _, applied = apply_completion_masking(
+        trainer, "LiquidAI/LFM2-8B-A1B", train_fn, detect_fn = detect
+    )
+    assert applied is True
+    assert seen == [inner]
diff --git a/studio/backend/tests/test_data_recipe_seed.py b/studio/backend/tests/test_data_recipe_seed.py
index 09e22116ed..58bbd24061 100644
--- a/studio/backend/tests/test_data_recipe_seed.py
+++ b/studio/backend/tests/test_data_recipe_seed.py
@@ -124,3 +124,100 @@ def test_unstructured_upload_import_errors_stay_generic(monkeypatch, tmp_path, e
     assert result.status == "error"
     assert result.error == "Text extraction failed."
     assert _block_files(seed_route) == []
+
+
+_TEST_UPLOAD_UID = "0f" * 16
+
+
+def test_remove_unstructured_block_deletes_directory(monkeypatch, tmp_path):
+    seed_route = _load_seed_route(monkeypatch, tmp_path)
+    _run_upload(seed_route, "notes.txt", b"hello", block_id = _TEST_UPLOAD_UID)
+    assert _block_files(seed_route, _TEST_UPLOAD_UID) != []
+
+    result = asyncio.run(seed_route.remove_unstructured_block(_TEST_UPLOAD_UID))
+
+    assert result == {"status": "ok", "deleted": True}
+    assert not (seed_route.UNSTRUCTURED_UPLOAD_ROOT / _TEST_UPLOAD_UID).exists()
+
+
+def test_remove_unstructured_block_missing_directory_is_ok(monkeypatch, tmp_path):
+    seed_route = _load_seed_route(monkeypatch, tmp_path)
+
+    result = asyncio.run(seed_route.remove_unstructured_block(_TEST_UPLOAD_UID))
+
+    assert result == {"status": "ok", "deleted": False}
+
+
+def test_remove_unstructured_block_rejects_unsafe_ids(monkeypatch, tmp_path):
+    seed_route = _load_seed_route(monkeypatch, tmp_path)
+
+    with pytest.raises(seed_route.HTTPException) as exc:
+        asyncio.run(seed_route.remove_unstructured_block("../escape"))
+
+    assert exc.value.status_code == 400
+
+
+def test_remove_unstructured_block_rejects_legacy_node_ids(monkeypatch, tmp_path):
+    seed_route = _load_seed_route(monkeypatch, tmp_path)
+    _run_upload(seed_route, "notes.txt", b"hello", block_id = "n1")
+    assert _block_files(seed_route, "n1") != []
+
+    with pytest.raises(seed_route.HTTPException) as exc:
+        asyncio.run(seed_route.remove_unstructured_block("n1"))
+
+    assert exc.value.status_code == 400
+    assert _block_files(seed_route, "n1") != []
+
+
+def test_remove_unstructured_block_rejects_symlink_escape(monkeypatch, tmp_path):
+    seed_route = _load_seed_route(monkeypatch, tmp_path)
+    outside = tmp_path / "outside"
+    outside.mkdir()
+    (outside / "victim.txt").write_text("keep me")
+    root = seed_route.UNSTRUCTURED_UPLOAD_ROOT
+    root.mkdir(parents = True)
+    (root / _TEST_UPLOAD_UID).symlink_to(outside)
+
+    with pytest.raises(seed_route.HTTPException) as exc:
+        asyncio.run(seed_route.remove_unstructured_block(_TEST_UPLOAD_UID))
+
+    assert exc.value.status_code == 400
+    assert (outside / "victim.txt").exists()
+
+
+def test_remove_unstructured_block_fails_if_directory_remains(monkeypatch, tmp_path):
+    seed_route = _load_seed_route(monkeypatch, tmp_path)
+    root = seed_route.UNSTRUCTURED_UPLOAD_ROOT
+    block_dir = root / _TEST_UPLOAD_UID
+    block_dir.mkdir(parents = True)
+    (block_dir / "victim.txt").write_text("keep me")
+
+    calls = []
+
+    def noop_rmtree(path, *args, **kwargs):
+        calls.append((path, args, kwargs))
+
+    monkeypatch.setattr(seed_route.shutil, "rmtree", noop_rmtree)
+
+    with pytest.raises(seed_route.HTTPException) as exc:
+        asyncio.run(seed_route.remove_unstructured_block(_TEST_UPLOAD_UID))
+
+    assert calls
+    assert exc.value.status_code == 500
+    assert block_dir.exists()
+
+
+def test_total_upload_quota_is_scoped_per_block(monkeypatch, tmp_path):
+    seed_route = _load_seed_route(monkeypatch, tmp_path)
+    monkeypatch.setattr(seed_route, "UNSTRUCTURED_RECIPE_UPLOAD_TOTAL_MAX_BYTES", 10)
+
+    first = _run_upload(seed_route, "a.txt", b"123456789")
+    assert first.status == "ok"
+
+    with pytest.raises(seed_route.HTTPException) as exc:
+        _run_upload(seed_route, "b.txt", b"123")
+    assert exc.value.status_code == 413
+
+    # Another block starts with its own untouched budget.
+    other = _run_upload(seed_route, "c.txt", b"123", block_id = "other")
+    assert other.status == "ok"
diff --git a/studio/backend/tests/test_hf_xet_fallback.py b/studio/backend/tests/test_hf_xet_fallback.py
index 4d73213d15..2fff744b64 100644
--- a/studio/backend/tests/test_hf_xet_fallback.py
+++ b/studio/backend/tests/test_hf_xet_fallback.py
@@ -287,7 +287,9 @@ def test_degrades_when_shared_helper_import_raises_importerror():
 
 def test_retries_under_light_gpu_init_when_import_fails(monkeypatch):
     """GPU detection in unsloth_zoo's __init__ raises NotImplementedError on a GPU-less host. The shim
-    retries under UNSLOTH_ZOO_DISABLE_GPU_INIT=1, restores the env, and degrades if the retry fails."""
+    retries under UNSLOTH_ZOO_DISABLE_GPU_INIT=1, restores the env, and degrades if the retry fails.
+    The backend loads lazily (first use of a heavy helper), so this triggers the load explicitly
+    before asserting the retry/degrade behavior."""
     import importlib
     import os
 
@@ -321,11 +323,15 @@ def test_retries_under_light_gpu_init_when_import_fails(monkeypatch):
     sys.meta_path.insert(0, finder)
     try:
         degraded = importlib.import_module("utils.hf_xet_fallback")
-        # First attempt without the light env, then a retry with it set.
+        # Import is light (lazy backend); unsloth_zoo not loaded yet.
+        assert seen_env == [], seen_env
+        # First use of a heavy helper triggers the load (attempt without the light env, then a retry
+        # with it set); accessing DownloadStallError drives it via __getattr__.
+        stall_error = degraded.DownloadStallError
         assert seen_env == [None, "1"], seen_env
         # Both attempts raised -> Studio still boots in degraded mode.
-        assert issubclass(degraded.DownloadStallError, RuntimeError)
-        # The env override must not leak past the import.
+        assert issubclass(stall_error, RuntimeError)
+        # The env override must not leak past the load.
         assert os.environ.get("UNSLOTH_ZOO_DISABLE_GPU_INIT") is None
     finally:
         sys.meta_path.remove(finder)
@@ -333,3 +339,31 @@ def test_retries_under_light_gpu_init_when_import_fails(monkeypatch):
         sys.modules.update(saved)
         if saved_shim is not None:
             sys.modules["utils.hf_xet_fallback"] = saved_shim
+
+
+def test_importing_child_should_disable_xet_stays_light(monkeypatch):
+    """Regression guard for the stale-transformers-sidecar bug: importing the shim (and
+    ``child_should_disable_xet``) must NOT pull in ``transformers``/``unsloth_zoo``. The worker calls
+    this at startup to decide the Xet env flip BEFORE activating the sidecar; an eager import here
+    would cache the default transformers 4.57.x in sys.modules, defeating the sidecar sys.path prepend
+    and breaking 5.x models (Qwen3.5/GLM/gemma-4)."""
+    import importlib
+
+    for name in [
+        m
+        for m in list(sys.modules)
+        if m == "transformers"
+        or m.startswith("transformers.")
+        or m == "unsloth_zoo"
+        or m.startswith("unsloth_zoo.")
+        or m == "utils.hf_xet_fallback"
+    ]:
+        monkeypatch.delitem(sys.modules, name, raising = False)
+
+    mod = importlib.import_module("utils.hf_xet_fallback")
+    # The lightweight decision works without the heavy backend.
+    assert mod.child_should_disable_xet({"disable_xet": True}) is True
+    assert mod.child_should_disable_xet({}) is False
+    # And nothing heavy was imported as a side effect.
+    assert "transformers" not in sys.modules, "importing the shim must not import transformers"
+    assert "unsloth_zoo" not in sys.modules, "importing the shim must not import unsloth_zoo"
diff --git a/studio/backend/tests/test_inference_dispatcher_resilience.py b/studio/backend/tests/test_inference_dispatcher_resilience.py
index 60d7503485..6184496d78 100644
--- a/studio/backend/tests/test_inference_dispatcher_resilience.py
+++ b/studio/backend/tests/test_inference_dispatcher_resilience.py
@@ -112,7 +112,7 @@ def test_route_llama_streaming_async_clients_disable_proxy_env():
             continue
         calls.append(node)
 
-    assert len(calls) == 4
+    assert len(calls) == 5
     for call in calls:
         assert any(
             kw.arg == "trust_env" and isinstance(kw.value, ast.Constant) and kw.value.value is False
diff --git a/studio/backend/tests/test_install_resolve_prebuilt.py b/studio/backend/tests/test_install_resolve_prebuilt.py
index e9941d9e62..e97ca47717 100644
--- a/studio/backend/tests/test_install_resolve_prebuilt.py
+++ b/studio/backend/tests/test_install_resolve_prebuilt.py
@@ -1,7 +1,8 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""install_llama_prebuilt.py: host->repo mapping and the --resolve-prebuilt mode.
+"""install_llama_prebuilt.py: the --resolve-prebuilt probe (plans against the fork
+by default; --published-repo overrides).
 
 These back the in-app update for source-build (markerless) installs: the backend
 asks the installer whether an official prebuilt exists for this host without
@@ -24,9 +25,7 @@ if str(_studio) not in sys.path:
 
 ilp = importlib.import_module("install_llama_prebuilt")
 
-if not hasattr(ilp, "published_repo_for_host") or not hasattr(
-    ilp, "resolve_simple_install_release_plans"
-):
+if not hasattr(ilp, "resolve_simple_install_release_plans"):
     pytest.skip("PR symbols not present - check branch", allow_module_level = True)
 
 FORK = ilp.DEFAULT_PUBLISHED_REPO  # unslothai/llama.cpp
@@ -56,71 +55,25 @@ def _host(**kw):
     return ilp.HostInfo(**base)
 
 
-def test_published_repo_for_host():
-    # CPU-only Linux (x64 and arm64) -> ggml-org upstream.
-    assert ilp.published_repo_for_host(_host(is_linux = True, is_x86_64 = True)) == UPSTREAM
-    assert (
-        ilp.published_repo_for_host(_host(is_linux = True, is_arm64 = True, machine = "aarch64"))
-        == UPSTREAM
-    )
-    # GPU Linux -> fork.
-    assert (
-        ilp.published_repo_for_host(_host(is_linux = True, is_x86_64 = True, has_usable_nvidia = True))
-        == FORK
-    )
-    assert ilp.published_repo_for_host(_host(is_linux = True, is_x86_64 = True, has_rocm = True)) == FORK
-    # CPU-only Windows -> ggml-org (setup.ps1: the fork ships no win-cpu bundle).
-    assert (
-        ilp.published_repo_for_host(_host(system = "Windows", is_windows = True, is_x86_64 = True))
-        == UPSTREAM
-    )
-    # GPU Windows -> fork.
-    assert (
-        ilp.published_repo_for_host(
-            _host(system = "Windows", is_windows = True, is_x86_64 = True, has_usable_nvidia = True)
-        )
-        == FORK
-    )
-    # macOS -> fork regardless of GPU (ggml-org macOS bundles need too-new macOS).
-    assert (
-        ilp.published_repo_for_host(
-            _host(system = "Darwin", is_macos = True, is_arm64 = True, machine = "arm64")
-        )
-        == FORK
-    )
-    # Linux with AMD tooling but no probed GPU -> fork (setup.sh routes on tooling).
-    assert (
-        ilp.published_repo_for_host(
-            _host(is_linux = True, is_x86_64 = True), linux_amd_tooling_present = True
-        )
-        == FORK
-    )
-    # The tooling hint is Linux-only: Windows CPU stays on ggml-org.
-    assert (
-        ilp.published_repo_for_host(
-            _host(system = "Windows", is_windows = True, is_x86_64 = True),
-            linux_amd_tooling_present = True,
-        )
-        == UPSTREAM
-    )
-
-
-def test_macos_intel_and_arm_both_route_to_fork():
-    # macOS uses the unslothai fork's own Mac prebuilts for BOTH arm64 and Intel;
-    # there is no longer any upstream-on-macOS default path, so the obsolete
-    # pre-macOS-26 pin (b9415) is gone.
-    assert (
-        ilp.published_repo_for_host(
-            _host(system = "Darwin", is_macos = True, is_arm64 = True, machine = "arm64")
-        )
-        == FORK
-    )
-    assert (
-        ilp.published_repo_for_host(
-            _host(system = "Darwin", is_macos = True, is_x86_64 = True, machine = "x86_64")
-        )
-        == FORK
+def test_force_cpu_clears_all_gpu_attributes_including_intel():
+    # --cpu-fallback is the "select the CPU prebuilt even when a GPU is present"
+    # escape hatch. It must drop EVERY GPU attribute, including has_intel_gpu, or
+    # the planner still prepends the Vulkan asset on an Intel-GPU host.
+    host = _host(
+        is_linux = True,
+        is_x86_64 = True,
+        has_usable_nvidia = True,
+        has_physical_nvidia = True,
+        has_rocm = True,
+        rocm_gfx_target = "gfx1100",
+        has_intel_gpu = True,
     )
+    forced = ilp._apply_host_overrides(host, force_cpu = True)
+    assert forced.has_usable_nvidia is False
+    assert forced.has_physical_nvidia is False
+    assert forced.has_rocm is False
+    assert forced.rocm_gfx_target is None
+    assert forced.has_intel_gpu is False
 
 
 def test_macos_upstream_pin_only_for_explicit_pre26_upstream():
@@ -188,15 +141,13 @@ def test_resolve_prebuilt_unavailable(monkeypatch, capsys):
     assert out["repo"] == FORK
 
 
-def test_resolve_prebuilt_linux_amd_tooling_routes_to_fork(monkeypatch, capsys):
-    # CPU-probed Linux host but rocminfo on PATH: the dispatch must route to the
-    # fork so a HIP source build is not offered an upstream CPU prebuilt.
-    monkeypatch.setattr(ilp, "detect_host", lambda: _host(is_linux = True, is_x86_64 = True))
-    monkeypatch.setattr(ilp.shutil, "which", lambda tool: tool == "rocminfo")
+def _run_resolve_capture_host(monkeypatch, capsys):
+    """Drive --resolve-prebuilt and return the host the resolver was handed."""
     seen = {}
 
     def _resolver(tag, host, repo, published_release_tag):
         seen["repo"] = repo
+        seen["host"] = host
         raise ilp.PrebuiltFallback("no asset")
 
     monkeypatch.setattr(ilp, "resolve_simple_install_release_plans", _resolver)
@@ -207,10 +158,33 @@ def test_resolve_prebuilt_linux_amd_tooling_routes_to_fork(monkeypatch, capsys):
     )
     assert ilp.main() == ilp.EXIT_SUCCESS
     out = json.loads(capsys.readouterr().out.strip().splitlines()[-1])
+    return seen, out
+
+
+def test_resolve_prebuilt_cpu_linux_routes_to_fork(monkeypatch, capsys):
+    # CPU-only Linux host (no GPU): the dispatch routes to the fork, which now
+    # ships the CPU prebuilt -- it no longer falls back to ggml-org upstream.
+    monkeypatch.setattr(ilp, "detect_host", lambda: _host(is_linux = True, is_x86_64 = True))
+    seen, out = _run_resolve_capture_host(monkeypatch, capsys)
     assert seen["repo"] == FORK
     assert out["repo"] == FORK
 
 
+def test_resolve_prebuilt_rocm_sdk_only_host_still_offered_cpu(monkeypatch, capsys):
+    # A CPU-only host that merely has ROCm/HIP SDK tools on PATH (no AMD GPU, so
+    # detect_host leaves has_rocm False) is a valid CPU-prebuilt target. The probe
+    # must NOT reclassify it as ROCm from tool presence alone and suppress the CPU
+    # bundle -- that would deny the fork CPU prebuilt to a legitimate CPU source
+    # build. The host is left CPU-only and resolves against the fork.
+    monkeypatch.setattr(ilp, "detect_host", lambda: _host(is_linux = True, is_x86_64 = True))
+    monkeypatch.setattr(
+        ilp.shutil, "which", lambda tool: "/opt/rocm/bin/hipconfig" if tool == "hipconfig" else None
+    )
+    seen, out = _run_resolve_capture_host(monkeypatch, capsys)
+    assert seen["repo"] == FORK
+    assert seen["host"].has_rocm is False
+
+
 # Blackwell floor is sm_100 (data-center B100/B200, B300/GB300), below consumer
 # sm_120 -- 120 wrongly excluded data-center hosts from the prebuilt selection.
 
@@ -360,3 +334,386 @@ def test_sm103_host_drops_cuda128_windows_build():
     )
     kept_b200 = ilp._drop_blackwell_incapable_windows_cuda(b200, [cuda128, cuda129])
     assert [a.name for a in kept_b200] == [cuda128.name, cuda129.name]
+
+
+def _upstream_release(tag, asset_names):
+    return {
+        "tag_name": tag,
+        "assets": [
+            {"name": n, "browser_download_url": f"https://example/{n}"} for n in asset_names
+        ],
+    }
+
+
+def test_direct_upstream_arm64_intel_prefers_vulkan():
+    # Auto-detected Intel GPU on Linux arm64 -> Vulkan prebuilt first, CPU
+    # second (mirrors the x86_64 branch; ggml-org ships the arm64 Vulkan asset).
+    host = _host(is_linux = True, is_arm64 = True, machine = "aarch64", has_intel_gpu = True)
+    rel = _upstream_release(
+        "b9925",
+        ["llama-b9925-bin-ubuntu-vulkan-arm64.tar.gz", "llama-b9925-bin-ubuntu-arm64.tar.gz"],
+    )
+    plan = ilp.direct_upstream_release_plan(rel, host, UPSTREAM, "latest")
+    kinds = [a.install_kind for a in plan.attempts]
+    assert kinds[0] == "linux-vulkan", kinds
+    assert "linux-arm64" in kinds
+    assert plan.attempts[0].name == "llama-b9925-bin-ubuntu-vulkan-arm64.tar.gz"
+
+
+def test_direct_upstream_intel_with_hidden_nvidia_is_cpu_only():
+    # A host with a physical NVIDIA hidden via CUDA_VISIBLE_DEVICES (physical
+    # True, usable False) + an Intel iGPU must NOT get the Vulkan archive even
+    # when planning directly against upstream: Vulkan ignores CUDA_VISIBLE_DEVICES
+    # and could grab the reserved card. It falls through to the CPU asset.
+    host = _host(
+        is_linux = True,
+        is_x86_64 = True,
+        has_intel_gpu = True,
+        has_physical_nvidia = True,
+        has_usable_nvidia = False,
+    )
+    rel = _upstream_release(
+        "b9925",
+        ["llama-b9925-bin-ubuntu-vulkan-x64.tar.gz", "llama-b9925-bin-ubuntu-x64.tar.gz"],
+    )
+    plan = ilp.direct_upstream_release_plan(rel, host, UPSTREAM, "latest")
+    assert [a.install_kind for a in plan.attempts] == ["linux-cpu"]
+
+
+def test_direct_upstream_arm64_without_intel_is_cpu_only():
+    host = _host(is_linux = True, is_arm64 = True, machine = "aarch64")
+    rel = _upstream_release(
+        "b9925",
+        ["llama-b9925-bin-ubuntu-vulkan-arm64.tar.gz", "llama-b9925-bin-ubuntu-arm64.tar.gz"],
+    )
+    plan = ilp.direct_upstream_release_plan(rel, host, UPSTREAM, "latest")
+    assert [a.install_kind for a in plan.attempts] == ["linux-arm64"]
+
+
+def test_direct_upstream_x86_intel_prefers_vulkan():
+    host = _host(is_linux = True, is_x86_64 = True, has_intel_gpu = True)
+    rel = _upstream_release(
+        "b9925",
+        ["llama-b9925-bin-ubuntu-vulkan-x64.tar.gz", "llama-b9925-bin-ubuntu-x64.tar.gz"],
+    )
+    plan = ilp.direct_upstream_release_plan(rel, host, UPSTREAM, "latest")
+    kinds = [a.install_kind for a in plan.attempts]
+    assert kinds[0] == "linux-vulkan", kinds
+    assert "linux-cpu" in kinds
+
+
+def test_linux_vulkan_health_glob_matches_bare_cpu_lib():
+    # The widened glob must cover both arch-suffixed (x64) and bare (arm64) CPU
+    # libs so a valid Vulkan install is not re-flagged unhealthy every check.
+    choice = ilp.AssetChoice(
+        repo = UPSTREAM,
+        tag = "b9925",
+        name = "llama-b9925-bin-ubuntu-vulkan-arm64.tar.gz",
+        url = "https://example/x",
+        source_label = "upstream",
+        install_kind = "linux-vulkan",
+    )
+    groups = ilp.runtime_payload_health_groups(choice)
+    assert ["libggml-cpu*.so*"] in groups
+    assert ["libggml-cpu-*.so*"] not in groups
+
+
+def test_route_to_vulkan_prebuilt_auto_intel_goes_upstream_and_drops_fork_pin():
+    # Routing fork -> upstream also drops the fork release pin, which is in a
+    # different tag namespace and would make the upstream resolver miss.
+    host = _host(is_linux = True, is_x86_64 = True, has_intel_gpu = True)
+    routed, repo, tag = ilp._route_to_vulkan_prebuilt(host, FORK, "b9596-mix-abc", force_cpu = False)
+    assert repo == UPSTREAM
+    assert tag == ""
+    assert routed.has_intel_gpu is True
+
+
+def test_route_to_vulkan_prebuilt_preserves_explicit_upstream_pin():
+    # A pin set WITH an explicit upstream repo is already on upstream -> kept.
+    host = _host(is_linux = True, is_x86_64 = True, has_intel_gpu = True)
+    _routed, repo, tag = ilp._route_to_vulkan_prebuilt(host, UPSTREAM, "b9596", force_cpu = False)
+    assert repo == UPSTREAM
+    assert tag == "b9596"
+
+
+def test_route_to_vulkan_prebuilt_cpu_fallback_wins():
+    # --cpu-fallback suppresses Vulkan routing even for an Intel host.
+    host = _host(is_linux = True, is_x86_64 = True, has_intel_gpu = True)
+    routed, repo, tag = ilp._route_to_vulkan_prebuilt(host, FORK, "b9596-mix-abc", force_cpu = True)
+    assert repo == FORK
+    assert tag == "b9596-mix-abc"
+    assert routed is host
+
+
+def test_route_to_vulkan_prebuilt_hidden_nvidia_not_rerouted():
+    # A mixed NVIDIA+Intel host that hid NVIDIA (CUDA_VISIBLE_DEVICES=""/-1):
+    # physical NVIDIA present but not usable. Must NOT auto-route to Vulkan, or
+    # Vulkan (which ignores CUDA_VISIBLE_DEVICES) could grab the reserved GPU.
+    host = _host(
+        is_linux = True,
+        is_x86_64 = True,
+        has_intel_gpu = True,
+        has_physical_nvidia = True,
+        has_usable_nvidia = False,
+    )
+    _routed, repo, _tag = ilp._route_to_vulkan_prebuilt(host, FORK, "", force_cpu = False)
+    assert repo == FORK
+
+
+def test_route_to_vulkan_prebuilt_rocm_host_not_rerouted():
+    # An Intel iGPU alongside a usable ROCm GPU stays on its ROCm/fork path.
+    host = _host(is_linux = True, is_x86_64 = True, has_intel_gpu = True, has_rocm = True)
+    _routed, repo, _tag = ilp._route_to_vulkan_prebuilt(host, FORK, "", force_cpu = False)
+    assert repo == FORK
+
+
+def test_route_to_vulkan_prebuilt_non_intel_unchanged():
+    host = _host(is_linux = True, is_x86_64 = True)
+    routed, repo, _tag = ilp._route_to_vulkan_prebuilt(host, FORK, "", force_cpu = False)
+    assert repo == FORK
+    assert routed is host
+
+
+def test_resolve_prebuilt_intel_host_routes_to_upstream(monkeypatch, capsys):
+    # The --resolve-prebuilt probe must agree with the install path: an
+    # auto-detected Intel host resolves against upstream (Vulkan), not the fork.
+    monkeypatch.setattr(
+        ilp, "detect_host", lambda: _host(is_linux = True, is_x86_64 = True, has_intel_gpu = True)
+    )
+    seen, out = _run_resolve_capture_host(monkeypatch, capsys)
+    assert seen["repo"] == UPSTREAM
+    assert out["repo"] == UPSTREAM
+
+
+# ---------------------------------------------------------------------------
+# windows_intel_gpu_in_registry: the in-process Windows Intel probe. A fake
+# winreg module stands in for the real registry so the walk runs anywhere.
+# ---------------------------------------------------------------------------
+
+
+class _FakeRegKey:
+    def __init__(
+        self,
+        subkeys = None,
+        values = None,
+        denied = False,
+    ):
+        self.subkeys = subkeys or {}
+        self.values = values or {}
+        self.denied = denied
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, *exc):
+        return False
+
+
+class _FakeWinreg:
+    HKEY_LOCAL_MACHINE = object()
+
+    def __init__(self, root_key):
+        self._root_key = root_key
+
+    def OpenKey(self, parent, name):
+        if parent is self.HKEY_LOCAL_MACHINE:
+            # Pin the production constant: a typo'd class GUID must fail here,
+            # not silently return the fake tree.
+            if name != ilp._WINDOWS_DISPLAY_CLASS_KEY:
+                raise FileNotFoundError(name)
+            if self._root_key is None:
+                raise FileNotFoundError(name)
+            return self._root_key
+        key = parent.subkeys.get(name)
+        if key is None:
+            # Real winreg raises OSError, never KeyError, for a missing key.
+            raise FileNotFoundError(name)
+        if key.denied:
+            raise PermissionError(name)
+        return key
+
+    def QueryInfoKey(self, key):
+        return (len(key.subkeys), len(key.values), 0)
+
+    def EnumKey(self, key, index):
+        return list(key.subkeys)[index]
+
+    def QueryValueEx(self, key, value_name):
+        if value_name not in key.values:
+            raise FileNotFoundError(value_name)
+        return (key.values[value_name], 1)
+
+
+def _probe_with_display_class(monkeypatch, adapters):
+    # The helper lazily does `import winreg`; plant the fake in sys.modules the
+    # same way unsloth_cli/tests/test_start.py fakes it for _refresh_windows_path.
+    monkeypatch.setitem(sys.modules, "winreg", _FakeWinreg(_FakeRegKey(subkeys = adapters)))
+    return ilp.windows_intel_gpu_in_registry()
+
+
+def test_windows_intel_registry_matches_vendor_id(monkeypatch):
+    assert (
+        _probe_with_display_class(
+            monkeypatch,
+            {
+                "0000": _FakeRegKey(
+                    values = {
+                        "MatchingDeviceId": r"PCI\VEN_8086&DEV_56A0&SUBSYS_12345678",
+                        "DriverDesc": "Intel(R) Arc(TM) A770 Graphics",
+                    }
+                ),
+            },
+        )
+        is True
+    )
+
+
+def test_windows_intel_registry_matches_driver_desc_without_device_id(monkeypatch):
+    assert (
+        _probe_with_display_class(
+            monkeypatch,
+            {
+                "0000": _FakeRegKey(values = {"DriverDesc": "Intel(R) UHD Graphics 630"}),
+            },
+        )
+        is True
+    )
+
+
+def test_windows_intel_registry_ignores_non_intel_adapters(monkeypatch):
+    assert (
+        _probe_with_display_class(
+            monkeypatch,
+            {
+                "0000": _FakeRegKey(
+                    values = {
+                        "MatchingDeviceId": r"PCI\VEN_10DE&DEV_2684",
+                        "DriverDesc": "NVIDIA GeForce RTX 4090",
+                    }
+                ),
+                "0001": _FakeRegKey(
+                    values = {
+                        "MatchingDeviceId": r"PCI\VEN_1002&DEV_744C",
+                        "DriverDesc": "AMD Radeon RX 7900 XTX",
+                    }
+                ),
+            },
+        )
+        is False
+    )
+
+
+def test_windows_intel_registry_skips_restricted_properties_subkey(monkeypatch):
+    # The real class key carries an ACL-restricted "Properties" subkey and can
+    # deny access to individual adapter keys; neither may abort the walk.
+    assert (
+        _probe_with_display_class(
+            monkeypatch,
+            {
+                "Properties": _FakeRegKey(denied = True),
+                "0000": _FakeRegKey(denied = True),
+                "0001": _FakeRegKey(
+                    values = {
+                        "MatchingDeviceId": r"PCI\VEN_8086&DEV_56A0",
+                    }
+                ),
+            },
+        )
+        is True
+    )
+
+
+def test_windows_intel_registry_missing_class_key_is_false(monkeypatch):
+    monkeypatch.setitem(sys.modules, "winreg", _FakeWinreg(None))
+    assert ilp.windows_intel_gpu_in_registry() is False
+
+
+def _detect_windows_host(
+    monkeypatch,
+    winreg_fake,
+    powershell_stdout = "",
+):
+    """Drive the real detect_host() as a GPU-less Windows host with a fake
+    registry, recording every run_capture invocation. Pins the wiring the
+    unit tests above cannot see: registry-first, CIM only on a registry miss."""
+    monkeypatch.setitem(sys.modules, "winreg", winreg_fake)
+    monkeypatch.setattr(ilp.platform, "system", lambda: "Windows")
+    monkeypatch.setattr(ilp.platform, "machine", lambda: "AMD64")
+    for _env in (
+        "CUDA_VISIBLE_DEVICES",
+        "HIP_VISIBLE_DEVICES",
+        "ROCR_VISIBLE_DEVICES",
+        "HIP_PATH",
+        "ROCM_PATH",
+    ):
+        monkeypatch.delenv(_env, raising = False)
+    monkeypatch.setattr(
+        ilp.shutil,
+        "which",
+        lambda name: "powershell" if name in ("powershell", "pwsh") else None,
+    )
+    captured = []
+
+    def _fake_run_capture(command, **kwargs):
+        captured.append(command[0])
+        if command[0] == "powershell":
+            return SimpleNamespace(returncode = 0, stdout = powershell_stdout, stderr = "")
+        return SimpleNamespace(returncode = 1, stdout = "", stderr = "")
+
+    monkeypatch.setattr(ilp, "run_capture", _fake_run_capture)
+    return ilp.detect_host(), captured
+
+
+def test_detect_host_registry_intel_skips_cim_probe(monkeypatch):
+    winreg = _FakeWinreg(
+        _FakeRegKey(
+            subkeys = {
+                "0000": _FakeRegKey(values = {"MatchingDeviceId": r"PCI\VEN_8086&DEV_56A0"}),
+            }
+        )
+    )
+    host, captured = _detect_windows_host(monkeypatch, winreg)
+    assert host.has_intel_gpu is True
+    assert "powershell" not in captured
+
+
+def test_detect_host_cim_fallback_fires_on_registry_miss(monkeypatch):
+    winreg = _FakeWinreg(
+        _FakeRegKey(
+            subkeys = {
+                "0000": _FakeRegKey(values = {"MatchingDeviceId": r"PCI\VEN_10DE&DEV_2684"}),
+            }
+        )
+    )
+    host, captured = _detect_windows_host(
+        monkeypatch, winreg, powershell_stdout = "Intel(R) Arc(TM) A770 Graphics"
+    )
+    assert host.has_intel_gpu is True
+    assert "powershell" in captured
+
+
+def test_windows_intel_registry_unexpected_error_is_false(monkeypatch):
+    # The probe is advisory: even a non-OSError bug in the walk must return
+    # False (deferring to the CIM fallback), never crash detect_host.
+    class _ExplodingWinreg:
+        HKEY_LOCAL_MACHINE = object()
+
+        def OpenKey(self, parent, name):
+            raise TypeError(name)
+
+    monkeypatch.setitem(sys.modules, "winreg", _ExplodingWinreg())
+    assert ilp.windows_intel_gpu_in_registry() is False
+
+
+def test_detect_host_cim_rescues_exploding_registry(monkeypatch):
+    class _ExplodingWinreg:
+        HKEY_LOCAL_MACHINE = object()
+
+        def OpenKey(self, parent, name):
+            raise TypeError(name)
+
+    host, captured = _detect_windows_host(
+        monkeypatch, _ExplodingWinreg(), powershell_stdout = "Intel(R) Arc(TM) A770 Graphics"
+    )
+    assert host.has_intel_gpu is True
+    assert "powershell" in captured
diff --git a/studio/backend/tests/test_llama_admission.py b/studio/backend/tests/test_llama_admission.py
new file mode 100644
index 0000000000..2f04e81926
--- /dev/null
+++ b/studio/backend/tests/test_llama_admission.py
@@ -0,0 +1,320 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+import asyncio
+import os
+import sys
+import threading
+
+import pytest
+
+_backend = os.path.join(os.path.dirname(__file__), "..")
+sys.path.insert(0, _backend)
+
+from core.inference import llama_admission
+from core.inference.llama_admission import (
+    ADMISSION_CONTROL_ENV,
+    ADMISSION_KEEPALIVE_INTERVAL_ENV,
+    ADMISSION_MAX_QUEUE_ENV,
+    ADMISSION_QUEUE_TIMEOUT_ENV,
+    DEFAULT_ADMISSION_KEEPALIVE_INTERVAL_S,
+    DEFAULT_ADMISSION_MAX_QUEUE,
+    DEFAULT_ADMISSION_QUEUE_TIMEOUT_S,
+    LlamaAdmissionConfig,
+    LlamaAdmissionQueueFull,
+    get_llama_admission_queue,
+    llama_admission_config_from_env,
+    reset_llama_admission_queues,
+)
+
+
+@pytest.fixture(autouse = True)
+def _reset_queues():
+    reset_llama_admission_queues()
+    yield
+    reset_llama_admission_queues()
+
+
+def test_admission_config_defaults(monkeypatch):
+    for name in (
+        ADMISSION_CONTROL_ENV,
+        ADMISSION_QUEUE_TIMEOUT_ENV,
+        ADMISSION_KEEPALIVE_INTERVAL_ENV,
+        ADMISSION_MAX_QUEUE_ENV,
+    ):
+        monkeypatch.delenv(name, raising = False)
+
+    config = llama_admission_config_from_env()
+
+    assert config.enabled is True
+    assert config.queue_timeout_s == DEFAULT_ADMISSION_QUEUE_TIMEOUT_S
+    assert config.keepalive_interval_s == DEFAULT_ADMISSION_KEEPALIVE_INTERVAL_S
+    assert config.max_queue == DEFAULT_ADMISSION_MAX_QUEUE
+
+
+def test_admission_config_env_overrides(monkeypatch):
+    monkeypatch.setenv(ADMISSION_CONTROL_ENV, "off")
+    monkeypatch.setenv(ADMISSION_QUEUE_TIMEOUT_ENV, "0")
+    monkeypatch.setenv(ADMISSION_KEEPALIVE_INTERVAL_ENV, "0.25")
+    monkeypatch.setenv(ADMISSION_MAX_QUEUE_ENV, "0")
+
+    config = llama_admission_config_from_env()
+
+    assert config.enabled is False
+    assert config.queue_timeout_s is None
+    assert config.keepalive_interval_s == 0.25
+    assert config.max_queue is None
+
+
+def test_admission_config_positive_queue_timeout_env(monkeypatch):
+    monkeypatch.setenv(ADMISSION_QUEUE_TIMEOUT_ENV, "600")
+
+    config = llama_admission_config_from_env()
+
+    assert config.queue_timeout_s == 600.0
+
+
+def test_fifo_capacity_one_grants_next_waiter_on_release():
+    async def _run():
+        queue = get_llama_admission_queue("http://llama.test")
+        config = LlamaAdmissionConfig()
+
+        first = queue.reserve(capacity = 1, config = config)
+        second = queue.reserve(capacity = 1, config = config)
+        third = queue.reserve(capacity = 1, config = config)
+
+        first_lease = first.lease_nowait()
+        assert first_lease is not None
+        assert second.lease_nowait() is None
+        assert third.lease_nowait() is None
+        assert queue.snapshot().queued == 2
+
+        first_lease.release()
+        second_lease = await second.wait(0.1)
+        assert second_lease is not None
+        assert third.lease_nowait() is None
+
+        second_lease.release()
+        third_lease = await third.wait(0.1)
+        assert third_lease is not None
+        third_lease.release()
+
+        snapshot = queue.snapshot()
+        assert snapshot.active == 0
+        assert snapshot.queued == 0
+
+    asyncio.run(_run())
+
+
+def test_queue_full_rejects_excess_waiter():
+    async def _run():
+        queue = get_llama_admission_queue("http://llama.test")
+        config = LlamaAdmissionConfig(max_queue = 1)
+
+        first = queue.reserve(capacity = 1, config = config)
+        queued = queue.reserve(capacity = 1, config = config)
+
+        assert first.lease_nowait() is not None
+        assert queued.lease_nowait() is None
+        with pytest.raises(LlamaAdmissionQueueFull):
+            queue.reserve(capacity = 1, config = config)
+
+    asyncio.run(_run())
+
+
+def test_disabled_admission_bypasses_active_slot_limit():
+    async def _run():
+        queue = get_llama_admission_queue("http://llama.test")
+        config = LlamaAdmissionConfig(enabled = False)
+
+        first = queue.reserve(capacity = 1, config = config)
+        second = queue.reserve(capacity = 1, config = config)
+
+        assert first.lease_nowait() is not None
+        assert second.lease_nowait() is not None
+        assert queue.snapshot().active == 0
+        assert queue.snapshot().queued == 0
+
+    asyncio.run(_run())
+
+
+def test_cancelling_promoted_waiter_releases_slot():
+    async def _run():
+        queue = get_llama_admission_queue("http://llama.test")
+        config = LlamaAdmissionConfig()
+
+        first = queue.reserve(capacity = 1, config = config)
+        second = queue.reserve(capacity = 1, config = config)
+        first_lease = first.lease_nowait()
+
+        first_lease.release()
+        await asyncio.sleep(0)
+        second.cancel()
+
+        snapshot = queue.snapshot()
+        assert snapshot.active == 0
+        assert snapshot.queued == 0
+
+    asyncio.run(_run())
+
+
+def test_cancelling_promoted_waiter_before_delivery_releases_slot():
+    async def _run():
+        queue = get_llama_admission_queue("http://llama.test")
+        config = LlamaAdmissionConfig()
+
+        first = queue.reserve(capacity = 1, config = config)
+        second = queue.reserve(capacity = 1, config = config)
+        first_lease = first.lease_nowait()
+
+        first_lease.release()
+        second.cancel()
+
+        snapshot = queue.snapshot()
+        assert snapshot.active == 0
+        assert snapshot.queued == 0
+
+    asyncio.run(_run())
+
+
+def test_external_waiter_future_cancel_invalidates_reservation():
+    async def _run():
+        queue = get_llama_admission_queue("http://llama.test")
+        config = LlamaAdmissionConfig()
+
+        first = queue.reserve(capacity = 1, config = config)
+        second = queue.reserve(capacity = 1, config = config)
+        first_lease = first.lease_nowait()
+        assert first_lease is not None
+        assert second._waiter is not None
+
+        second._waiter.future.cancel()
+
+        assert second.lease_nowait() is None
+        assert second.is_cancelled is True
+        assert await second.wait(0.01) is None
+
+        first_lease.release()
+        snapshot = queue.snapshot()
+        assert snapshot.active == 0
+        assert snapshot.queued == 0
+
+    asyncio.run(_run())
+
+
+def test_wait_returns_none_when_waiter_future_cancelled_during_wait():
+    async def _run():
+        queue = get_llama_admission_queue("http://llama.test")
+        config = LlamaAdmissionConfig()
+
+        first = queue.reserve(capacity = 1, config = config)
+        second = queue.reserve(capacity = 1, config = config)
+        first_lease = first.lease_nowait()
+        assert first_lease is not None
+        assert second._waiter is not None
+
+        wait_task = asyncio.create_task(second.wait(1.0))
+        await asyncio.sleep(0)
+        second._waiter.future.cancel()
+
+        assert await asyncio.wait_for(wait_task, timeout = 0.1) is None
+        assert second.is_cancelled is True
+
+        first_lease.release()
+        snapshot = queue.snapshot()
+        assert snapshot.active == 0
+        assert snapshot.queued == 0
+
+    asyncio.run(_run())
+
+
+def test_capacity_increase_promotes_existing_waiter_fifo():
+    async def _run():
+        queue = get_llama_admission_queue("http://llama.test")
+        config = LlamaAdmissionConfig()
+
+        first = queue.reserve(capacity = 1, config = config)
+        second = queue.reserve(capacity = 1, config = config)
+
+        first_lease = first.lease_nowait()
+        assert first_lease is not None
+        assert second.lease_nowait() is None
+        assert queue.snapshot().active == 1
+        assert queue.snapshot().queued == 1
+
+        third = queue.reserve(capacity = 2, config = config)
+
+        second_lease = await second.wait(0.1)
+        assert second_lease is not None
+        assert third.lease_nowait() is None
+
+        snapshot = queue.snapshot()
+        assert snapshot.capacity == 2
+        assert snapshot.active == 2
+        assert snapshot.queued == 1
+
+        first_lease.release()
+        third_lease = await third.wait(0.1)
+        assert third_lease is not None
+
+        second_lease.release()
+        third_lease.release()
+        snapshot = queue.snapshot()
+        assert snapshot.active == 0
+        assert snapshot.queued == 0
+
+    asyncio.run(_run())
+
+
+def test_lease_release_is_idempotent_under_concurrent_calls():
+    async def _run():
+        queue = get_llama_admission_queue("http://llama.test")
+        config = LlamaAdmissionConfig()
+
+        reservation = queue.reserve(capacity = 1, config = config)
+        lease = reservation.lease_nowait()
+        assert lease is not None
+
+        threads = [threading.Thread(target = lease.release) for _ in range(16)]
+        for thread in threads:
+            thread.start()
+        for thread in threads:
+            thread.join()
+
+        snapshot = queue.snapshot()
+        assert snapshot.active == 0
+        assert snapshot.queued == 0
+
+    asyncio.run(_run())
+
+
+def test_new_key_evicts_idle_prior_load_queues():
+    # Each model load carries a fresh ephemeral port, so a new base_url key must
+    # not leave the drained queues from earlier loads accumulating forever.
+    get_llama_admission_queue("http://127.0.0.1:1001")
+    get_llama_admission_queue("http://127.0.0.1:1002")
+    assert set(llama_admission._QUEUES) == {"http://127.0.0.1:1002"}
+
+    get_llama_admission_queue("http://127.0.0.1:1003")
+    assert set(llama_admission._QUEUES) == {"http://127.0.0.1:1003"}
+
+
+def test_new_key_retains_in_flight_prior_load_queue():
+    config = LlamaAdmissionConfig()
+    busy = get_llama_admission_queue("http://127.0.0.1:2001")
+
+    async def _run():
+        reservation = busy.reserve(capacity = 1, config = config)
+        lease = reservation.lease_nowait()
+        assert lease is not None
+
+        # A new load must not drop a queue that still has an in-flight request.
+        get_llama_admission_queue("http://127.0.0.1:2002")
+        assert set(llama_admission._QUEUES) == {"http://127.0.0.1:2001", "http://127.0.0.1:2002"}
+
+        # Once it drains, the next load reclaims it.
+        lease.release()
+        get_llama_admission_queue("http://127.0.0.1:2003")
+        assert set(llama_admission._QUEUES) == {"http://127.0.0.1:2003"}
+
+    asyncio.run(_run())
diff --git a/studio/backend/tests/test_llama_cpp_effective_parallel_slots.py b/studio/backend/tests/test_llama_cpp_effective_parallel_slots.py
new file mode 100644
index 0000000000..5525bc3ea9
--- /dev/null
+++ b/studio/backend/tests/test_llama_cpp_effective_parallel_slots.py
@@ -0,0 +1,53 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+import os
+import sys
+
+import pytest
+
+_backend = os.path.join(os.path.dirname(__file__), "..")
+sys.path.insert(0, _backend)
+
+from core.inference import llama_cpp as llama_cpp_module
+from core.inference.llama_cpp import LlamaCppBackend
+
+
+@pytest.fixture
+def backend(monkeypatch):
+    monkeypatch.setattr(LlamaCppBackend, "_kill_orphaned_servers", lambda self: 0)
+    monkeypatch.setattr(llama_cpp_module.atexit, "register", lambda *_args, **_kwargs: None)
+    return LlamaCppBackend()
+
+
+def test_effective_parallel_slots_initial_value_is_one(backend):
+    assert backend.effective_parallel_slots == 1
+
+
+def test_effective_parallel_slots_commit_uses_final_positive_parallel(backend):
+    backend._commit_effective_parallel_slots(3)
+
+    assert backend.effective_parallel_slots == 3
+
+
+@pytest.mark.parametrize("value", [None, 0, -2, "not-an-int"])
+def test_effective_parallel_slots_commit_invalid_value_falls_back_to_one(backend, value):
+    backend._commit_effective_parallel_slots(value)
+
+    assert backend.effective_parallel_slots == 1
+
+
+def test_effective_parallel_slots_reset_returns_to_one(backend):
+    backend._commit_effective_parallel_slots(4)
+
+    backend._reset_effective_parallel_slots()
+
+    assert backend.effective_parallel_slots == 1
+
+
+def test_effective_parallel_slots_unload_resets_to_one(backend):
+    backend._commit_effective_parallel_slots(4)
+
+    backend.unload_model()
+
+    assert backend.effective_parallel_slots == 1
diff --git a/studio/backend/tests/test_llama_cpp_freshness.py b/studio/backend/tests/test_llama_cpp_freshness.py
index 2a2e113585..08e1334ac9 100644
--- a/studio/backend/tests/test_llama_cpp_freshness.py
+++ b/studio/backend/tests/test_llama_cpp_freshness.py
@@ -137,9 +137,9 @@ def test_read_install_marker_finds_windows_cmake_layout(tmp_path):
 
 @pytest.mark.parametrize("repo", ["unslothai/llama.cpp", "ggml-org/llama.cpp"])
 def test_read_install_marker_carries_published_repo_dynamically(tmp_path, repo):
-    # The freshness check queries whichever release repo the marker records,
-    # so CUDA (unslothai), CPU/macOS (ggml-org), and ROCm all get the right
-    # "latest" tag.
+    # The freshness check queries whichever release repo the marker records:
+    # new installs record the fork, legacy CPU/macOS markers still say ggml-org,
+    # and both must get the right "latest" tag.
     install_dir = tmp_path / "llama.cpp"
     _write_marker(install_dir, tag = "b9000", published_repo = repo)
     bin_path = _fake_binary(install_dir, layout = "cmake")
diff --git a/studio/backend/tests/test_llama_cpp_stream_cancel.py b/studio/backend/tests/test_llama_cpp_stream_cancel.py
new file mode 100644
index 0000000000..1c73f4d17c
--- /dev/null
+++ b/studio/backend/tests/test_llama_cpp_stream_cancel.py
@@ -0,0 +1,81 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+import contextlib
+import os
+import sys
+import threading
+
+import httpx
+import pytest
+
+_backend = os.path.join(os.path.dirname(__file__), "..")
+sys.path.insert(0, _backend)
+
+from core.inference.llama_cpp import LlamaCppBackend, _LlamaStreamCancelled
+
+
+def _backend_stub() -> LlamaCppBackend:
+    backend = LlamaCppBackend.__new__(LlamaCppBackend)
+    backend._process = object()
+    backend._healthy = True
+    backend._port = 48848
+    backend._effective_context_length = 4096
+    backend._supports_reasoning = False
+    backend._reasoning_always_on = False
+    backend._reasoning_style = "enable_thinking"
+    backend._supports_preserve_thinking = False
+    return backend
+
+
+def test_stream_cancel_uses_internal_exception_not_generator_exit():
+    class FakeResponse:
+        status_code = 200
+
+        def close(self):
+            pass
+
+    class FakeStream:
+        def __enter__(self):
+            return FakeResponse()
+
+        def __exit__(self, *_args):
+            return False
+
+    class FakeClient:
+        def stream(self, *_args, **_kwargs):
+            return FakeStream()
+
+    cancel_event = threading.Event()
+
+    with pytest.raises(Exception) as exc_info:
+        with LlamaCppBackend._stream_with_retry(
+            FakeClient(),
+            "http://llama.test/v1/chat/completions",
+            {},
+            cancel_event,
+        ):
+            cancel_event.set()
+            raise httpx.ReadError("client closed")
+
+    assert exc_info.type is _LlamaStreamCancelled
+    assert not issubclass(exc_info.type, GeneratorExit)
+
+
+def test_generate_chat_completion_swallows_internal_stream_cancel(monkeypatch):
+    backend = _backend_stub()
+
+    @contextlib.contextmanager
+    def fake_open_stream(*_args, **_kwargs):
+        raise _LlamaStreamCancelled
+
+    monkeypatch.setattr(backend, "_open_stream", fake_open_stream)
+
+    chunks = list(
+        backend.generate_chat_completion(
+            [{"role": "user", "content": "hi"}],
+            cancel_event = threading.Event(),
+        )
+    )
+
+    assert chunks == []
diff --git a/studio/backend/tests/test_llama_cpp_update.py b/studio/backend/tests/test_llama_cpp_update.py
index 4ceffbf75b..f405ebcbd1 100644
--- a/studio/backend/tests/test_llama_cpp_update.py
+++ b/studio/backend/tests/test_llama_cpp_update.py
@@ -448,6 +448,48 @@ def test_start_update_happy_path(monkeypatch, tmp_path):
     assert popen_kwargs["env"]["UNSLOTH_PROGRESS_PERCENT_STEP"] == "5"
 
 
+def test_start_update_preserves_vulkan_via_env(monkeypatch, tmp_path):
+    # A Vulkan install (marker asset carries 'vulkan') must re-assert
+    # UNSLOTH_FORCE_VULKAN on update, or detect_host on a GPU box re-routes to
+    # CUDA/ROCm and silently replaces the Vulkan build.
+    install_dir = tmp_path / "llama.cpp"
+    binary = _write_install(
+        install_dir,
+        "b9493",
+        repo = "ggml-org/llama.cpp",
+        asset = "llama-b9493-bin-ubuntu-vulkan-x64.tar.gz",
+    )
+    monkeypatch.setattr(upd, "_find_binary", lambda: binary)
+    monkeypatch.setattr(upd, "_installer_script", lambda: tmp_path / "install_llama_prebuilt.py")
+    monkeypatch.setattr(freshness, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: "b9518")
+
+    def _on_start(cmd):
+        _write_install(
+            install_dir,
+            "b9518",
+            repo = "ggml-org/llama.cpp",
+            asset = "llama-b9518-bin-ubuntu-vulkan-x64.tar.gz",
+        )
+
+    popen_kwargs: dict = {}
+    _patch_installer_popen(
+        monkeypatch,
+        lines = ["installed\n"],
+        on_start = _on_start,
+        captured_kwargs = popen_kwargs,
+    )
+
+    assert upd.start_update()["started"] is True
+    deadline = time.time() + 10
+    while time.time() < deadline:
+        job = upd.get_update_status()["job"]
+        if job["state"] in ("success", "error"):
+            break
+        time.sleep(0.05)
+    assert job["state"] == "success", job
+    assert popen_kwargs["env"]["UNSLOTH_FORCE_VULKAN"] == "1"
+
+
 def test_start_update_reports_full_release_tag(monkeypatch, tmp_path):
     install_dir = tmp_path / "llama.cpp"
     binary = _write_install(install_dir, "b9595")
@@ -594,9 +636,10 @@ def test_install_cmd_fork_rocm_marker_forwards_has_rocm(monkeypatch, tmp_path):
 
 
 def test_install_cmd_ggml_cpu_marker_has_no_cpu_fallback(monkeypatch, tmp_path):
-    # CPU installs come from ggml-org. Re-running into the same install-dir/repo
-    # reproduces the same CPU bundle; --cpu-fallback (which force-drops GPU
-    # detection) is reserved for setup.sh's arm64 rescue and must not appear here.
+    # Legacy CPU installs recorded a ggml-org marker (new installs use the fork).
+    # Re-running into the same install-dir/repo reproduces the same CPU bundle;
+    # --cpu-fallback (which force-drops GPU detection) is reserved for setup.sh's
+    # arm64 rescue and must not appear here.
     cmd = _capture_install_cmd(
         monkeypatch,
         tmp_path,
diff --git a/studio/backend/tests/test_llama_cpp_vulkan_probe.py b/studio/backend/tests/test_llama_cpp_vulkan_probe.py
new file mode 100644
index 0000000000..92aaab4873
--- /dev/null
+++ b/studio/backend/tests/test_llama_cpp_vulkan_probe.py
@@ -0,0 +1,193 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Vulkan free-VRAM reader regression tests on a synthetic probe output.
+
+Covers the post-probe handling in
+``LlamaCppBackend._get_gpu_free_memory_vulkan``:
+
+  * integrated GPUs (probe reports is_igpu=1) leave a flat per-device host
+    margin matching llama.cpp's --fit-target, so context auto-sizing can't
+    over-commit shared RAM, and report total 0 (shared RAM is not a budget),
+  * discrete GPUs (is_igpu=0) keep their free untouched and pass their real
+    total through so the fit can reserve absolute headroom,
+  * an inherited ``GGML_VK_VISIBLE_DEVICES`` is passed through to ggml unchanged
+    (ggml applies it), not stripped or filtered in Python -- the probe reports
+    ggml's compact ordinal, which load_model pins with ``--device Vulkan<i>``.
+
+The ggml Vulkan library is never loaded: subprocess.run is mocked to emit
+the tab-separated lines the real ``_vulkan_probe.py`` would print.
+"""
+
+from __future__ import annotations
+
+import subprocess
+import sys
+import types as _types
+from pathlib import Path
+from unittest import mock
+
+import pytest
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+import importlib as _importlib  # noqa: E402
+
+
+def _maybe_stub(name: str, builder):
+    try:
+        _importlib.import_module(name)
+    except ImportError:
+        sys.modules[name] = builder()
+
+
+def _build_loggers_stub():
+    m = _types.ModuleType("loggers")
+    m.get_logger = lambda name: __import__("logging").getLogger(name)
+    return m
+
+
+_maybe_stub("loggers", _build_loggers_stub)
+_maybe_stub("structlog", lambda: _types.ModuleType("structlog"))
+
+from core.inference import llama_cpp as _llama_mod  # noqa: E402
+from core.inference.llama_cpp import (  # noqa: E402
+    LlamaCppBackend,
+    _llama_lib_dir,
+    _vulkan_lib_filename,
+)
+
+MIB = 1024 * 1024
+GIB = 1024 * MIB
+
+
+def _make_vulkan_install(tmp_path: Path) -> str:
+    """A binary whose sibling dir holds the Vulkan ggml lib, so the
+    reader's ``is_vulkan_backend`` sibling-file check passes."""
+    bindir = tmp_path / "build" / "bin"
+    bindir.mkdir(parents = True)
+    binary = bindir / ("llama-server.exe" if sys.platform == "win32" else "llama-server")
+    binary.write_bytes(b"stub")
+    (bindir / _vulkan_lib_filename()).write_bytes(b"stub")
+    return str(binary)
+
+
+def _mock_probe(rows: list[str], captured_env: dict | None = None):
+    """Patch subprocess.run so the _vulkan_probe.py call returns ``rows``
+    (already tab-formatted), recording the env it was launched with."""
+    real_run = subprocess.run
+
+    def fake_run(cmd, *args, **kwargs):
+        if isinstance(cmd, list) and any("_vulkan_probe" in str(c) for c in cmd):
+            if captured_env is not None:
+                captured_env.clear()
+                captured_env.update(kwargs.get("env") or {})
+            return subprocess.CompletedProcess(
+                args = cmd, returncode = 0, stdout = "\n".join(rows), stderr = ""
+            )
+        return real_run(cmd, *args, **kwargs)
+
+    return mock.patch("subprocess.run", side_effect = fake_run)
+
+
+def _row(
+    idx: int,
+    free_bytes: int,
+    is_igpu: int,
+    total_bytes: int = 0,
+) -> str:
+    return f"{idx}\t{free_bytes}\t{is_igpu}\t{total_bytes}"
+
+
+def test_integrated_gpu_leaves_host_margin(tmp_path):
+    binary = _make_vulkan_install(tmp_path)
+    # iGPU with 30 GiB free; reserve a flat 1024 MiB (llama.cpp --fit-target).
+    # total stays 0: shared system RAM is not a VRAM budget for the fit.
+    rows = [_row(0, 30 * GIB, is_igpu = 1, total_bytes = 32 * GIB)]
+    with _mock_probe(rows):
+        gpus = LlamaCppBackend._get_gpu_free_memory_vulkan(binary)
+    assert gpus == [(0, 30 * 1024 - 1024, 0)], gpus
+
+
+def test_discrete_gpu_free_is_untouched_and_total_passed_through(tmp_path):
+    binary = _make_vulkan_install(tmp_path)
+    # 6 GiB free on a partially occupied 24 GiB card: free is untouched and the
+    # real total flows through so the fit reserves absolute headroom (CUDA/ROCm
+    # parity) instead of the looser free*frac budget.
+    rows = [_row(0, 6 * GIB, is_igpu = 0, total_bytes = 24 * GIB)]
+    with _mock_probe(rows):
+        gpus = LlamaCppBackend._get_gpu_free_memory_vulkan(binary)
+    assert gpus == [(0, 6 * 1024, 24 * 1024)], gpus
+
+
+def test_large_discrete_gpu_is_untouched(tmp_path):
+    binary = _make_vulkan_install(tmp_path)
+    # A 48 GiB discrete card stays untouched regardless of size; only the
+    # iGPU flag triggers the host margin, never a VRAM/RAM ratio.
+    rows = [_row(0, 47 * GIB, is_igpu = 0, total_bytes = 48 * GIB)]
+    with _mock_probe(rows):
+        gpus = LlamaCppBackend._get_gpu_free_memory_vulkan(binary)
+    assert gpus == [(0, 47 * 1024, 48 * 1024)], gpus
+
+
+def test_inherited_visible_devices_mask_is_passed_through_to_probe(tmp_path, monkeypatch):
+    # The mask is NOT stripped or filtered in Python: ggml parses it in raw
+    # physical-device space while this probe reports the compact post-filter
+    # ordinal, so mixing spaces would be wrong. It is passed through unchanged
+    # so ggml applies it to the same device list the launch will enumerate.
+    binary = _make_vulkan_install(tmp_path)
+    monkeypatch.setenv("GGML_VK_VISIBLE_DEVICES", "1")
+    captured: dict = {}
+    rows = [_row(0, 23 * GIB, is_igpu = 0, total_bytes = 24 * GIB)]
+    with _mock_probe(rows, captured_env = captured):
+        LlamaCppBackend._get_gpu_free_memory_vulkan(binary)
+    assert captured.get("GGML_VK_VISIBLE_DEVICES") == "1", captured
+
+
+def test_vulkan_pin_args_uses_device_names_not_env_mask():
+    # Pin by compact device name via --device (the space the probe reports and
+    # the registry names), never by writing a compact ordinal into the raw
+    # GGML_VK_VISIBLE_DEVICES index space.
+    assert LlamaCppBackend._vulkan_pin_args([0]) == ["--device", "Vulkan0"]
+    assert LlamaCppBackend._vulkan_pin_args([1, 2]) == ["--device", "Vulkan1,Vulkan2"]
+    assert LlamaCppBackend._vulkan_pin_args(None) == []
+    assert LlamaCppBackend._vulkan_pin_args([]) == []
+
+
+def test_vulkan_only_build_is_detected(tmp_path):
+    binary = _make_vulkan_install(tmp_path)
+    assert LlamaCppBackend._is_vulkan_backend(binary) is True
+
+
+def test_multi_backend_build_is_not_vulkan_only(tmp_path):
+    # A custom build that ships CUDA (or HIP) alongside Vulkan must NOT be
+    # treated as Vulkan-only, or its CUDA GPU would be probed/pinned as a Vulkan
+    # device; defer to the CUDA/HIP path instead.
+    binary = _make_vulkan_install(tmp_path)
+    cuda = "ggml-cuda.dll" if sys.platform == "win32" else "libggml-cuda.so"
+    (_llama_lib_dir(binary) / cuda).write_bytes(b"stub")
+    assert LlamaCppBackend._is_vulkan_backend(binary) is False
+
+
+@pytest.mark.skipif(sys.platform == "win32", reason = "shell wrapper fallback is POSIX")
+def test_shell_wrapper_entrypoint_resolves_to_real_lib_dir(tmp_path):
+    # create_exec_entrypoint falls back to a #!/bin/sh wrapper at the install root
+    # when it cannot symlink; _find_llama_server_binary returns that root entrypoint,
+    # so _llama_lib_dir must follow the wrapper's exec target to build/bin -- else
+    # _is_vulkan_backend misses libggml-vulkan.so and the Vulkan probe/pin silently
+    # never engage on a valid Vulkan install.
+    import os
+
+    binary = _make_vulkan_install(tmp_path)  # tmp_path/build/bin/llama-server + vulkan lib
+    bindir = Path(binary).parent
+    wrapper = tmp_path / "llama-server"
+    wrapper.write_text('#!/bin/sh\nexec "$(dirname "$0")/build/bin/llama-server" "$@"\n')
+    os.chmod(wrapper, 0o755)
+    assert _llama_lib_dir(str(wrapper)) == bindir
+    assert LlamaCppBackend._is_vulkan_backend(str(wrapper)) is True
+
+
+if __name__ == "__main__":
+    raise SystemExit(pytest.main([__file__, "-v"]))
diff --git a/studio/backend/tests/test_llama_route_timeouts.py b/studio/backend/tests/test_llama_route_timeouts.py
index 24866bd03e..b4666e3d18 100644
--- a/studio/backend/tests/test_llama_route_timeouts.py
+++ b/studio/backend/tests/test_llama_route_timeouts.py
@@ -203,3 +203,87 @@ def test_preheader_send_cleanup_on_disconnect_and_cancel():
 
     asyncio.run(_run(False))
     asyncio.run(_run(True))
+
+
+def test_stream_stall_timeout_callable_re_resolved_each_read():
+    # The OpenAI passthrough passes a callable so the stall bound can switch to
+    # the short post-terminal grace mid-stream; it must be re-resolved per read,
+    # not captured once at generator start.
+    async def _run():
+        response = SimpleNamespace(request = SimpleNamespace(extensions = {"timeout": {}}))
+        values = iter([100.0, 2.0])
+        seen = []
+
+        class _Request:
+            async def is_disconnected(self):
+                return False
+
+        class _Items:
+            def __init__(self):
+                self.count = 0
+
+            async def __anext__(self):
+                self.count += 1
+                if self.count > 3:
+                    raise StopAsyncIteration
+                return "data: {}"
+
+        async for _ in inf_mod._aiter_llama_stream_items(
+            _Items(),
+            cancel_event = threading.Event(),
+            request = _Request(),
+            response = response,
+            first_token_deadline = time.monotonic() + 1,
+            post_first_item_read_timeout_s = lambda: next(values, 5.0),
+        ):
+            seen.append(response.request.extensions["timeout"].get("read"))
+
+        assert len(seen) == 3
+        # The callable is resolved right after the first item (arming the
+        # post-first window) and again before each later read, consuming
+        # successive values.
+        assert seen[0] == 100.0
+        assert 1.0 <= seen[1] <= 2.0
+        assert 4.0 <= seen[2] <= 5.0
+
+    asyncio.run(_run())
+
+
+def test_stream_stall_timeout_disabled_clears_read_timeout():
+    # UNSLOTH_OPENAI_COMPAT_STREAM_STALL_TIMEOUT=0 disables the stall guard, so
+    # the callable returns None. Once a chunk has arrived the leftover
+    # first-token read timeout must be cleared, else a long post-first-chunk gap
+    # trips a stale deadline the operator asked to turn off.
+    async def _run():
+        response = SimpleNamespace(request = SimpleNamespace(extensions = {"timeout": {}}))
+        seen = []
+
+        class _Request:
+            async def is_disconnected(self):
+                return False
+
+        class _Items:
+            def __init__(self):
+                self.count = 0
+
+            async def __anext__(self):
+                self.count += 1
+                if self.count > 2:
+                    raise StopAsyncIteration
+                return "data: {}"
+
+        async for _ in inf_mod._aiter_llama_stream_items(
+            _Items(),
+            cancel_event = threading.Event(),
+            request = _Request(),
+            response = response,
+            first_token_deadline = time.monotonic() + 5,
+            post_first_item_read_timeout_s = lambda: None,
+        ):
+            seen.append(response.request.extensions["timeout"].get("read"))
+
+        # The first-token path armed a finite read timeout; after the first chunk
+        # with the guard disabled, it is cleared to None on every subsequent read.
+        assert seen == [None, None], seen
+
+    asyncio.run(_run())
diff --git a/studio/backend/tests/test_mlx_inference_backend.py b/studio/backend/tests/test_mlx_inference_backend.py
index ac4088fb25..55a3198a6b 100644
--- a/studio/backend/tests/test_mlx_inference_backend.py
+++ b/studio/backend/tests/test_mlx_inference_backend.py
@@ -4,6 +4,8 @@ import sys
 import types
 from types import SimpleNamespace
 
+import pytest
+
 
 class _DummyMetal:
     @staticmethod
@@ -185,6 +187,129 @@ def test_mlx_inference_vlm_lora_uses_unsloth_loader_without_native_adapter_rewri
     assert isinstance(backend._tokenizer, _DummyTokenizer)
 
 
+def test_mlx_inference_distributed_vlm_forwards_group_to_fast_mlx(monkeypatch):
+    _install_fake_mlx(monkeypatch)
+    calls = []
+    _install_fake_fast_mlx(monkeypatch, calls)
+    from core.inference.mlx_inference import MLXInferenceBackend
+
+    group = SimpleNamespace(size = lambda: 2, rank = lambda: 0)
+    config = SimpleNamespace(identifier = "fake/vlm", is_vision = True, is_lora = False)
+    for mode, group_key in (("tensor", "tensor_group"), ("pipeline", "pipeline_group")):
+        calls.clear()
+        assert MLXInferenceBackend().load_model(config, parallel_mode = mode, distributed_group = group)
+        _, kwargs = calls.pop()
+        assert kwargs["text_only"] is False and kwargs[group_key] is group
+
+    calls.clear()
+    singleton = SimpleNamespace(size = lambda: 1, rank = lambda: 0)
+    assert MLXInferenceBackend().load_model(
+        config, parallel_mode = "tensor", distributed_group = singleton
+    )
+    assert not {"tensor_group", "pipeline_group"} & set(calls.pop()[1])
+
+    config = SimpleNamespace(identifier = "fake/adapter", is_vision = False, is_lora = True)
+    with pytest.raises(ValueError, match = "LoRA adapter repos"):
+        MLXInferenceBackend().load_model(config, parallel_mode = "tensor", distributed_group = group)
+
+
+@pytest.mark.parametrize("accepts_backend", (True, False))
+def test_mlx_distributed_init_selects_jaccl_backend(monkeypatch, accepts_backend):
+    _install_fake_mlx(monkeypatch)
+    from core.inference.mlx_inference import _init_mlx_distributed
+
+    group = SimpleNamespace(rank = lambda: 1, size = lambda: 2)
+    calls = []
+
+    def _init(**kwargs):
+        calls.append(kwargs)
+        if kwargs and not accepts_backend:
+            raise TypeError("backend keyword unsupported")
+        return group
+
+    sys.modules["mlx.core"].distributed = SimpleNamespace(init = _init)
+    monkeypatch.setenv("MLX_JACCL_COORDINATOR", "127.0.0.1:12345")
+    monkeypatch.setenv("MLX_IBV_DEVICES", "/tmp/devices.json")
+
+    assert _init_mlx_distributed() == (group, 1, 2)
+    assert calls == ([{"backend": "jaccl"}] if accepts_backend else [{"backend": "jaccl"}, {}])
+
+
+def test_worker_share_object_receives_distributed_payload(monkeypatch):
+    from core.inference import worker
+
+    shared_obj = {"type": "turn", "text": "hi"}
+    payload = worker._encode_share_object(shared_obj)
+
+    def _array(value):
+        val = value.item() if hasattr(value, "item") else value
+        return SimpleNamespace(
+            item = lambda: val,
+            tolist = lambda: list(val) if hasattr(val, "__iter__") else [val],
+        )
+
+    mlx_pkg = types.ModuleType("mlx")
+    mlx_core = types.ModuleType("mlx.core")
+    mlx_core.uint8 = "uint8"
+    mlx_core.array = _array
+    mlx_core.zeros = lambda *_a, **_k: _array([])
+
+    def _all_sum(value, group = None):
+        value = value.item() if hasattr(value, "item") else value
+        return _array(len(payload)) if value == 0 else _array(payload)
+
+    mlx_core.distributed = SimpleNamespace(all_sum = _all_sum)
+    mlx_pkg.core = mlx_core
+    monkeypatch.setitem(sys.modules, "mlx", mlx_pkg)
+    monkeypatch.setitem(sys.modules, "mlx.core", mlx_core)
+
+    responses = []
+    worker._handle_share_object(
+        SimpleNamespace(
+            _distributed_group = object(),
+            _distributed_rank = 1,
+            _distributed_world_size = 2,
+        ),
+        {"type": "share_object", "request_id": "rid", "object": None},
+        SimpleNamespace(put = responses.append),
+    )
+
+    response = responses[0]
+    assert response["object"] == shared_obj
+
+
+def test_worker_share_object_oversize_notifies_peers(monkeypatch):
+    from core.inference import worker
+
+    calls = []
+
+    mlx_pkg = types.ModuleType("mlx")
+    mlx_core = types.ModuleType("mlx.core")
+    mlx_core.array = lambda value, **_kwargs: SimpleNamespace(item = lambda: value)
+    mlx_core.eval = lambda value: value
+    mlx_core.distributed = SimpleNamespace(
+        all_sum = lambda value, group = None: calls.append(value.item()) or value
+    )
+    mlx_pkg.core = mlx_core
+    monkeypatch.setitem(sys.modules, "mlx", mlx_pkg)
+    monkeypatch.setitem(sys.modules, "mlx.core", mlx_core)
+    monkeypatch.setattr(worker, "_SHARE_OBJECT_MAX_BYTES", 8)
+
+    responses = []
+    worker._handle_share_object(
+        SimpleNamespace(
+            _distributed_group = object(),
+            _distributed_rank = 0,
+            _distributed_world_size = 2,
+        ),
+        {"type": "share_object", "request_id": "rid", "object": {"text": "too long"}},
+        SimpleNamespace(put = responses.append),
+    )
+
+    assert calls == [worker._SHARE_OBJECT_ERROR_SIZE]
+    assert responses[0]["type"] == "share_error"
+
+
 # Regression: generate_chat_response must accept the four template kwargs
 # (tools / enable_thinking / reasoning_effort / preserve_thinking) so the route
 # layer can forward UI toggles. The old signature raised
diff --git a/studio/backend/tests/test_mtp_drafter_companion.py b/studio/backend/tests/test_mtp_drafter_companion.py
index d5e8d13652..86e528ae67 100644
--- a/studio/backend/tests/test_mtp_drafter_companion.py
+++ b/studio/backend/tests/test_mtp_drafter_companion.py
@@ -23,7 +23,11 @@ if _BACKEND_DIR not in sys.path:
 
 from hub.utils.download_manifest import ExpectedFile
 from hub.utils.gguf import is_mtp_drafter_path
-from hub.utils.gguf_plan import build_gguf_variant_plans, plan_from_expected_files
+from hub.utils.gguf_plan import (
+    build_gguf_variant_plans,
+    plan_from_expected_files,
+    preferred_mtp_sibling,
+)
 from utils.models.model_config import (
     _is_mtp_drafter,
     detect_gguf_model,
@@ -37,6 +41,8 @@ from utils.models.model_config import (
 DRAFTER_CASES = [
     ("mtp-gemma-4-12b-it.gguf", True),
     ("MTP/gemma-4-12b-it-Q8_0-MTP.gguf", True),
+    # New-scheme MTP/ copies carry the mtp- basename prefix too.
+    ("MTP/mtp-gemma-4-E4B-it-BF16.gguf", True),
     ("foo/MTP/bar.gguf", True),
     ("gemma-4-12b-it-Q8_0.gguf", False),
     # Baked-in Qwen MTP repos: the head is inside the main GGUF, the file
@@ -274,3 +280,178 @@ def test_detect_gguf_model_rejects_mtp_subdir_copy(tmp_path):
     assert detect_gguf_model(str(copy)) is None
     # Selecting the MTP dir itself must not surface the copies as models.
     assert detect_gguf_model(str(sub)) is None
+
+
+# ── Root drafter wins over new-scheme MTP/ copies ────────────────────
+# The MTP/ copies were renamed to share the mtp- basename prefix (e.g.
+# MTP/mtp-gemma-4-E4B-it-BF16.gguf). Auto-fetch/load must still resolve the
+# small repo-root drafter, not a sort-first MTP/ copy (uppercase precedes
+# lowercase, so the subdir path would otherwise win).
+
+NEW_SCHEME_SIBLINGS = [
+    _sib("gemma-4-12b-it-Q4_K_M.gguf", 4_000, "main-q4"),
+    _sib("gemma-4-12b-it-Q8_0.gguf", 8_000, "main-q8"),
+    _sib("mtp-gemma-4-12b-it.gguf", 100, "drafter"),
+    _sib("MTP/mtp-gemma-4-12b-it-Q8_0.gguf", 100, "mtp-sub-q8"),
+    _sib("MTP/mtp-gemma-4-12b-it-BF16.gguf", 200, "mtp-sub-bf16"),
+    _sib("mmproj-F16.gguf", 500, "mmproj"),
+]
+
+
+def test_preferred_mtp_sibling_prefers_root_over_new_scheme_copies():
+    picked = preferred_mtp_sibling(NEW_SCHEME_SIBLINGS)
+    assert picked is not None and picked.rfilename == "mtp-gemma-4-12b-it.gguf"
+
+
+def test_variant_plans_new_scheme_uses_root_drafter():
+    plans = build_gguf_variant_plans(NEW_SCHEME_SIBLINGS)
+    assert set(plans) == {"q4_k_m", "q8_0"}
+    for plan in plans.values():
+        assert "mtp-gemma-4-12b-it.gguf" in plan.target_filenames
+        assert not any("MTP/" in name for name in plan.target_filenames)
+        assert "drafter" in plan.companion_hashes
+    # Download size = main + mmproj + root drafter (not the 200-byte BF16 copy).
+    assert plans["q4_k_m"].download_size_bytes == 4_600
+
+
+def test_download_mtp_prefers_root_over_new_scheme_copies(monkeypatch):
+    # _pick_mtp is nested; capture it via the companion-download seam.
+    from core.inference.llama_cpp import LlamaCppBackend
+
+    monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)  # online: skip reuse probe
+    captured = {}
+
+    def _fake_companion(
+        *,
+        hf_repo,
+        hf_token,
+        pick,
+        label,
+        cancel_event = None,
+    ):
+        captured["pick"] = pick
+        return None
+
+    b = LlamaCppBackend()
+    b._download_companion_gguf = _fake_companion
+    b._download_mtp(hf_repo = "unsloth/gemma-4-E4B-it-qat-mobile-GGUF")
+
+    repo_files = [
+        "MTP/mtp-gemma-4-E4B-it-BF16.gguf",
+        "MTP/mtp-gemma-4-E4B-it-Q4_0.gguf",
+        "MTP/mtp-gemma-4-E4B-it-Q8_0.gguf",
+        "gemma-4-E4B-it-qat-UD-Q2_K_XL.gguf",
+        "mmproj-F16.gguf",
+        "mtp-gemma-4-E4B-it.gguf",
+    ]
+    assert captured["pick"](repo_files) == "mtp-gemma-4-E4B-it.gguf"
+
+
+# ── Reuse an on-disk drafter offline; fetch fresh online ─────────────
+
+
+def _seed_snapshot(tmp_path, names):
+    snap = tmp_path / "snap"
+    for rel in names:
+        f = snap / rel
+        f.parent.mkdir(parents = True, exist_ok = True)
+        f.write_bytes(b"x")
+    return snap
+
+
+def test_download_mtp_reuses_cached_root_drafter_offline(tmp_path, monkeypatch):
+    import utils.models.model_config as mc
+    from core.inference.llama_cpp import LlamaCppBackend
+
+    monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+    snap = _seed_snapshot(
+        tmp_path,
+        [
+            "gemma-4-E4B-it-qat-UD-Q2_K_XL.gguf",
+            "mtp-gemma-4-E4B-it.gguf",
+            "MTP/mtp-gemma-4-E4B-it-BF16.gguf",
+            "mmproj-F16.gguf",
+        ],
+    )
+    monkeypatch.setattr(mc, "_iter_hf_cache_snapshots", lambda repo: [snap])
+
+    got = LlamaCppBackend()._download_mtp(hf_repo = "unsloth/gemma-4-E4B-it-qat-mobile-GGUF")
+    assert got is not None and Path(got).name == "mtp-gemma-4-E4B-it.gguf"
+
+
+def test_download_mtp_reuses_cached_subdir_copy_when_no_root_offline(tmp_path, monkeypatch):
+    # Pre-fix build may have fetched only the MTP/ copy; reuse it offline.
+    import utils.models.model_config as mc
+    from core.inference.llama_cpp import LlamaCppBackend
+
+    monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+    snap = _seed_snapshot(
+        tmp_path,
+        [
+            "gemma-4-E4B-it-qat-UD-Q2_K_XL.gguf",
+            "MTP/mtp-gemma-4-E4B-it-BF16.gguf",
+        ],
+    )
+    monkeypatch.setattr(mc, "_iter_hf_cache_snapshots", lambda repo: [snap])
+
+    got = LlamaCppBackend()._download_mtp(hf_repo = "unsloth/gemma-4-E4B-it-qat-mobile-GGUF")
+    assert got is not None and Path(got).name == "mtp-gemma-4-E4B-it-BF16.gguf"
+
+
+def test_download_mtp_prefers_root_across_snapshots_offline(tmp_path, monkeypatch):
+    # A newer partial snapshot holds only the MTP/ copy; an older one has the
+    # root. Must still return the small root, not the large subdir copy.
+    import utils.models.model_config as mc
+    from core.inference.llama_cpp import LlamaCppBackend
+
+    monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+    snap_partial = _seed_snapshot(tmp_path / "new", ["MTP/mtp-gemma-4-E4B-it-BF16.gguf"])
+    snap_full = _seed_snapshot(tmp_path / "old", ["mtp-gemma-4-E4B-it.gguf"])
+    monkeypatch.setattr(mc, "_iter_hf_cache_snapshots", lambda repo: [snap_partial, snap_full])
+
+    got = LlamaCppBackend()._download_mtp(hf_repo = "unsloth/gemma-4-E4B-it-qat-mobile-GGUF")
+    assert got is not None and Path(got).name == "mtp-gemma-4-E4B-it.gguf"
+
+
+def test_download_mtp_reuse_follows_snapshot_order_offline(tmp_path, monkeypatch):
+    # Two snapshots both hold a root drafter; newest-first order must win so a
+    # fresh main GGUF is not paired with a stale drafter revision.
+    import utils.models.model_config as mc
+    from core.inference.llama_cpp import LlamaCppBackend
+
+    monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+    newest = _seed_snapshot(tmp_path / "newest", ["mtp-gemma-4-E4B-it.gguf"])
+    oldest = _seed_snapshot(tmp_path / "oldest", ["mtp-gemma-4-E4B-it.gguf"])
+    monkeypatch.setattr(mc, "_iter_hf_cache_snapshots", lambda repo: [newest, oldest])
+
+    got = LlamaCppBackend()._download_mtp(hf_repo = "unsloth/gemma-4-E4B-it-qat-mobile-GGUF")
+    assert got is not None and Path(got).parent.parent.name == "newest"
+
+
+def test_download_mtp_online_skips_cache_reuse(tmp_path, monkeypatch):
+    # Online, do not reuse a cached copy: go to the download path so a changed
+    # drafter is refetched (hf_hub_download checks the current revision).
+    import utils.models.model_config as mc
+    from core.inference.llama_cpp import LlamaCppBackend
+
+    monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+    snap = _seed_snapshot(tmp_path, ["mtp-gemma-4-E4B-it.gguf"])
+    monkeypatch.setattr(mc, "_iter_hf_cache_snapshots", lambda repo: [snap])
+
+    reached = {}
+
+    def _fake_companion(
+        *,
+        hf_repo,
+        hf_token,
+        pick,
+        label,
+        cancel_event = None,
+    ):
+        reached["hit"] = True
+        return None
+
+    b = LlamaCppBackend()
+    b._download_companion_gguf = _fake_companion
+    assert b._download_mtp(hf_repo = "unsloth/gemma-4-E4B-it-qat-mobile-GGUF") is None
+    assert reached.get("hit") is True
diff --git a/studio/backend/tests/test_offline_gguf_cache_fallback.py b/studio/backend/tests/test_offline_gguf_cache_fallback.py
index 4499881c4d..e24e2ca451 100644
--- a/studio/backend/tests/test_offline_gguf_cache_fallback.py
+++ b/studio/backend/tests/test_offline_gguf_cache_fallback.py
@@ -79,9 +79,11 @@ from huggingface_hub import constants as hf_constants
 
 from core.inference.llama_cpp import (
     LlamaCppBackend,
+    _cached_colocated_split_main,
     _gguf_files_for_variant,
     _hf_offline_if_dns_dead,
     _probe_dns_dead,
+    _resolve_repo_id_casing,
 )
 from utils.models.model_config import (
     _detect_gguf_from_hf_cache,
@@ -217,7 +219,7 @@ class TestGgufVariantFileResolution:
             downloaded.append(filename)
             return f"/fake/{repo_id}/{filename}"
 
-        monkeypatch.setenv("HF_HUB_CACHE", str(tmp_path))
+        monkeypatch.setattr(hf_constants, "HF_HUB_CACHE", str(tmp_path))
         with (
             patch(
                 "huggingface_hub.list_repo_files",
@@ -239,6 +241,214 @@ class TestGgufVariantFileResolution:
         assert downloaded == ["tinyllamas/stories260K.gguf"]
         assert out == "/fake/ggml-org/models/tinyllamas/stories260K.gguf"
 
+    def test_download_reuses_older_snapshot_when_current_ref_snapshot_is_partial(
+        self, monkeypatch, hf_cache
+    ):
+        # Cross-snapshot reuse is an offline-resilience path: online, hf_hub_download
+        # resumes the partial current-ref download and revalidates the revision instead
+        # of serving an older snapshot's same-name blob.
+        monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+        backend = LlamaCppBackend()
+        repo = "unsloth/vision-GGUF"
+        old = _build_cache(
+            hf_cache,
+            repo,
+            {"model-UD-Q4_K_XL.gguf": 4},
+            snapshot_sha = "a" * 40,
+        )
+        _build_cache(
+            hf_cache,
+            repo,
+            {"mtp-model.gguf": 1},
+            snapshot_sha = "b" * 40,
+        )
+
+        def fake_get_paths_info(
+            _repo_id,
+            paths,
+            token = None,
+        ):
+            return [_types.SimpleNamespace(path = path, size = 4) for path in paths if path]
+
+        def fail_download(*_args, **_kwargs):
+            raise AssertionError("should reuse the cached GGUF instead of downloading")
+
+        with (
+            patch(
+                "huggingface_hub.list_repo_files",
+                lambda *_a, **_k: ["model-UD-Q4_K_XL.gguf", "mtp-model.gguf"],
+            ),
+            patch("huggingface_hub.get_paths_info", fake_get_paths_info),
+            patch("huggingface_hub.try_to_load_from_cache", lambda *_a, **_k: None),
+            patch("core.inference.llama_cpp.hf_hub_download_with_xet_fallback", fail_download),
+        ):
+            out = backend._download_gguf(
+                hf_repo = repo,
+                hf_variant = "UD-Q4_K_XL",
+            )
+
+        assert out == str(old / "model-UD-Q4_K_XL.gguf")
+
+    def test_download_reuses_cached_gguf_when_lowercase_partial_cache_shadows_it(
+        self, monkeypatch, hf_cache
+    ):
+        # Case-variant cross-dir reuse is offline-only; online the canonical repo id
+        # resolves up front and hf_hub_download fetches the current revision.
+        monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+        backend = LlamaCppBackend()
+        canonical_repo = "unsloth/gemma-4-E2B-it-GGUF"
+        requested_repo = "unsloth/gemma-4-e2b-it-gguf"
+        gguf_file = "gemma-4-E2B-it-UD-Q4_K_XL.gguf"
+        snap = _build_cache(
+            hf_cache,
+            canonical_repo,
+            {gguf_file: 4},
+            snapshot_sha = "a" * 40,
+        )
+        lower_snap = _build_cache(
+            hf_cache,
+            requested_repo,
+            {"mtp-gemma-4-E2B-it.gguf": 1},
+            snapshot_sha = "b" * 40,
+        )
+        os.utime(lower_snap, (2000, 2000))
+        os.utime(snap, (1000, 1000))
+        seen_repos: list[str] = []
+
+        def fake_list_repo_files(repo_id, token = None):
+            seen_repos.append(repo_id)
+            return [gguf_file]
+
+        def fake_get_paths_info(
+            repo_id,
+            paths,
+            token = None,
+        ):
+            seen_repos.append(repo_id)
+            return [_types.SimpleNamespace(path = path, size = 4) for path in paths if path]
+
+        def fake_cache(repo_id, filename, *args, **kwargs):
+            seen_repos.append(repo_id)
+            return str(snap / filename) if repo_id == canonical_repo else None
+
+        def fail_download(*_args, **_kwargs):
+            raise AssertionError("should reuse the cached GGUF instead of downloading")
+
+        with (
+            patch("huggingface_hub.list_repo_files", fake_list_repo_files),
+            patch("huggingface_hub.get_paths_info", fake_get_paths_info),
+            patch("huggingface_hub.try_to_load_from_cache", fake_cache),
+            patch("core.inference.llama_cpp.hf_hub_download_with_xet_fallback", fail_download),
+        ):
+            out = backend._download_gguf(
+                hf_repo = requested_repo,
+                hf_variant = "UD-Q4_K_XL",
+            )
+
+        assert out == str(snap / gguf_file)
+        assert seen_repos
+
+    def test_download_online_does_not_reuse_old_snapshot(self, monkeypatch, hf_cache):
+        # Online, an older same-name snapshot must not be served (it may be a stale
+        # revision); hf_hub_download is called so the current revision is fetched and
+        # its etag revalidated.
+        monkeypatch.delenv("HF_HUB_OFFLINE", raising = False)
+        backend = LlamaCppBackend()
+        repo = "unsloth/vision-GGUF"
+        _build_cache(hf_cache, repo, {"model-UD-Q4_K_XL.gguf": 4}, snapshot_sha = "a" * 40)
+        downloaded: list[str] = []
+
+        def fake_get_paths_info(
+            _repo_id,
+            paths,
+            token = None,
+        ):
+            return [_types.SimpleNamespace(path = p, size = 4) for p in paths if p]
+
+        def fake_download(
+            repo_id,
+            filename,
+            token = None,
+            **kwargs,
+        ):
+            downloaded.append(filename)
+            return f"/fresh/{filename}"
+
+        with (
+            patch(
+                "huggingface_hub.list_repo_files",
+                lambda *_a, **_k: ["model-UD-Q4_K_XL.gguf"],
+            ),
+            patch("huggingface_hub.get_paths_info", fake_get_paths_info),
+            patch("huggingface_hub.try_to_load_from_cache", lambda *_a, **_k: None),
+            patch("core.inference.llama_cpp.hf_hub_download_with_xet_fallback", fake_download),
+        ):
+            out = backend._download_gguf(hf_repo = repo, hf_variant = "UD-Q4_K_XL")
+
+        assert downloaded == ["model-UD-Q4_K_XL.gguf"]
+        assert out == "/fresh/model-UD-Q4_K_XL.gguf"
+
+    def test_download_reuses_older_snapshot_when_offline_env_is_true(self, monkeypatch, hf_cache):
+        # HF_HUB_OFFLINE accepts truthy spellings beyond "1" (true/yes/on); the offline
+        # cache reuse must trigger for those too, otherwise the earlier Hub calls run
+        # offline while this branch still attempts hf_hub_download and the cached GGUF
+        # cannot load.
+        monkeypatch.setenv("HF_HUB_OFFLINE", "true")
+        backend = LlamaCppBackend()
+        repo = "unsloth/vision-GGUF"
+        old = _build_cache(hf_cache, repo, {"model-UD-Q4_K_XL.gguf": 4}, snapshot_sha = "a" * 40)
+
+        def fake_get_paths_info(
+            _repo_id,
+            paths,
+            token = None,
+        ):
+            return [_types.SimpleNamespace(path = p, size = 4) for p in paths if p]
+
+        def fail_download(*_args, **_kwargs):
+            raise AssertionError("should reuse the cached GGUF instead of downloading")
+
+        with (
+            patch("huggingface_hub.list_repo_files", lambda *_a, **_k: ["model-UD-Q4_K_XL.gguf"]),
+            patch("huggingface_hub.get_paths_info", fake_get_paths_info),
+            patch("huggingface_hub.try_to_load_from_cache", lambda *_a, **_k: None),
+            patch("core.inference.llama_cpp.hf_hub_download_with_xet_fallback", fail_download),
+        ):
+            out = backend._download_gguf(hf_repo = repo, hf_variant = "UD-Q4_K_XL")
+
+        assert out == str(old / "model-UD-Q4_K_XL.gguf")
+
+    def test_download_companion_resolves_from_case_variant_snapshot_offline(
+        self, monkeypatch, hf_cache
+    ):
+        # Offline, resolve_cached_repo_id_case can keep a partial lower-case spelling,
+        # so the companion (mmproj) must resolve from whichever case-variant snapshot
+        # actually holds it rather than being dropped by an hf_hub_download on the
+        # wrong casing.
+        monkeypatch.setenv("HF_HUB_OFFLINE", "1")
+        backend = LlamaCppBackend()
+        canonical_repo = "unsloth/gemma-4-E2B-it-GGUF"
+        requested_repo = "unsloth/gemma-4-e2b-it-gguf"
+        snap = _build_cache(hf_cache, canonical_repo, {"mmproj-F16.gguf": 4}, snapshot_sha = "a" * 40)
+        # A partial lower-case dir exists so casing resolution keeps the requested spelling.
+        _build_cache(hf_cache, requested_repo, {"config.json": 1}, snapshot_sha = "b" * 40)
+
+        _offline_exc = type("OfflineModeIsEnabled", (Exception,), {})
+
+        def fake_list_repo_files(repo_id, token = None):
+            raise _offline_exc("offline")
+
+        def fail_download(*_args, **_kwargs):
+            raise AssertionError("should resolve the companion from cache, not download")
+
+        with (
+            patch("huggingface_hub.list_repo_files", fake_list_repo_files),
+            patch("core.inference.llama_cpp.hf_hub_download_with_xet_fallback", fail_download),
+        ):
+            out = backend._download_mmproj(hf_repo = requested_repo)
+
+        assert out == str(snap / "mmproj-F16.gguf")
+
     def test_download_includes_uppercase_split_gguf_shards(self, monkeypatch, tmp_path):
         backend = LlamaCppBackend()
         downloaded: list[str] = []
@@ -264,7 +474,7 @@ class TestGgufVariantFileResolution:
             downloaded.append(filename)
             return f"/fake/{repo_id}/{filename}"
 
-        monkeypatch.setenv("HF_HUB_CACHE", str(tmp_path))
+        monkeypatch.setattr(hf_constants, "HF_HUB_CACHE", str(tmp_path))
         with (
             patch("huggingface_hub.list_repo_files", lambda *_a, **_k: files),
             patch("huggingface_hub.get_paths_info", fake_get_paths_info),
@@ -279,6 +489,48 @@ class TestGgufVariantFileResolution:
         assert downloaded == files
         assert out == "/fake/org/repo/model-Q4_K_M-00001-of-00002.GGUF"
 
+    def test_download_refetches_split_gguf_when_shards_span_snapshots(self, monkeypatch, hf_cache):
+        # The cached main shard lives in an older snapshot; its sibling shard is only
+        # in a newer, separate snapshot. Reusing the main shard alone would leave
+        # llama.cpp unable to resolve the sibling, so the whole set must be re-fetched
+        # together (co-located) rather than served split across snapshot dirs.
+        backend = LlamaCppBackend()
+        repo = "org/split"
+        files = [
+            "model-Q4_K_M-00001-of-00002.gguf",
+            "model-Q4_K_M-00002-of-00002.gguf",
+        ]
+        _build_cache(hf_cache, repo, {files[0]: 4}, snapshot_sha = "a" * 40)
+        _build_cache(hf_cache, repo, {files[1]: 4}, snapshot_sha = "b" * 40)
+        downloaded: list[str] = []
+
+        def fake_get_paths_info(
+            _repo_id,
+            paths,
+            token = None,
+        ):
+            return [_types.SimpleNamespace(path = p, size = 4) for p in paths if p]
+
+        def fake_download(
+            repo_id,
+            filename,
+            token = None,
+            **_kwargs,
+        ):
+            downloaded.append(filename)
+            return f"/fake/{repo_id}/{filename}"
+
+        with (
+            patch("huggingface_hub.list_repo_files", lambda *_a, **_k: files),
+            patch("huggingface_hub.get_paths_info", fake_get_paths_info),
+            patch("huggingface_hub.try_to_load_from_cache", lambda *_a, **_k: None),
+            patch("core.inference.llama_cpp.hf_hub_download_with_xet_fallback", fake_download),
+        ):
+            out = backend._download_gguf(hf_repo = repo, hf_variant = "Q4_K_M")
+
+        assert downloaded == files
+        assert out == f"/fake/{repo}/{files[0]}"
+
 
 def _siblings(items: dict[str, int]):
     """Mock ``hf_model_info(...).siblings`` payload."""
@@ -315,6 +567,21 @@ class TestIterHfCacheSnapshots:
         out = list(_iter_hf_cache_snapshots("unsloth/multi"))
         assert [p.name for p in out] == ["b" * 40, "a" * 40]
 
+    def test_skips_snapshot_when_mtime_is_unavailable(self, hf_cache, monkeypatch):
+        stale = _build_cache(hf_cache, "unsloth/multi", {"x.gguf": 1}, snapshot_sha = "a" * 40)
+        good = _build_cache(hf_cache, "unsloth/multi", {"y.gguf": 1}, snapshot_sha = "b" * 40)
+        original_stat = Path.stat
+
+        def flaky_stat(self, *args, **kwargs):
+            if self == stale:
+                raise FileNotFoundError(str(self))
+            return original_stat(self, *args, **kwargs)
+
+        monkeypatch.setattr(Path, "stat", flaky_stat)
+
+        out = list(_iter_hf_cache_snapshots("unsloth/multi"))
+        assert out == [good]
+
     def test_repo_id_match_is_case_insensitive(self, hf_cache):
         _build_cache(hf_cache, "unsloth/Foo-GGUF", {"Foo-Q4_K_M.gguf": 1})
         # Lookup with different org/name casing still resolves
@@ -347,6 +614,87 @@ class TestListGgufVariantsFromCache:
         assert _list_gguf_variants_from_hf_cache("unsloth/absent") is None
 
 
+class TestCachedColocatedSplitMain:
+    def test_prefers_older_complete_snapshot_over_newer_partial(self, hf_cache):
+        # Newer snapshot has only shard 1; older snapshot has the complete set. The
+        # complete older snapshot must win so the split GGUF can load co-located.
+        shard1 = "m-00001-of-00002.gguf"
+        shard2 = "m-00002-of-00002.gguf"
+        old = _build_cache(
+            hf_cache, "unsloth/split-GGUF", {shard1: 100, shard2: 100}, snapshot_sha = "a" * 40
+        )
+        new = _build_cache(hf_cache, "unsloth/split-GGUF", {shard1: 100}, snapshot_sha = "b" * 40)
+        os.utime(old, (1000, 1000))
+        os.utime(new, (2000, 2000))
+
+        main = _cached_colocated_split_main("unsloth/split-GGUF", shard1, [shard2], {})
+        assert main is not None
+        assert main.startswith(str(old))
+
+    def test_returns_none_when_shards_span_snapshots(self, hf_cache):
+        shard1 = "m-00001-of-00002.gguf"
+        shard2 = "m-00002-of-00002.gguf"
+        a = _build_cache(hf_cache, "unsloth/split-GGUF", {shard1: 100}, snapshot_sha = "a" * 40)
+        b = _build_cache(hf_cache, "unsloth/split-GGUF", {shard2: 100}, snapshot_sha = "b" * 40)
+        os.utime(a, (1000, 1000))
+        os.utime(b, (2000, 2000))
+
+        assert _cached_colocated_split_main("unsloth/split-GGUF", shard1, [shard2], {}) is None
+
+
+class TestResolveRepoIdCasing:
+    def test_maps_to_canonical_casing(self, monkeypatch):
+        monkeypatch.setattr(
+            "utils.paths.resolve_cached_repo_id_case",
+            lambda repo: "unsloth/Gemma-4-GGUF" if repo.lower() == "unsloth/gemma-4-gguf" else repo,
+        )
+        # A companion download passed the resolved id reads the same cache entry
+        # as the main GGUF instead of missing it under the requested casing.
+        assert _resolve_repo_id_casing("unsloth/gemma-4-gguf") == "unsloth/Gemma-4-GGUF"
+
+    def test_passthrough_on_resolver_error(self, monkeypatch):
+        def boom(_repo):
+            raise RuntimeError("resolver unavailable")
+
+        monkeypatch.setattr("utils.paths.resolve_cached_repo_id_case", boom)
+        assert _resolve_repo_id_casing("unsloth/gemma-4-gguf") == "unsloth/gemma-4-gguf"
+
+    def test_companion_only_newer_snapshot_does_not_shadow_real_variants(self, hf_cache):
+        # A newer snapshot holds only a vision projector fetched on demand,
+        # while the quant files live in an older snapshot. The newer snapshot
+        # must not shadow the real variants; the vision flag carries over.
+        old = _build_cache(
+            hf_cache,
+            "unsloth/vision-GGUF",
+            {"vision-Q4_K_M.gguf": 100},
+            snapshot_sha = "a" * 40,
+        )
+        new = _build_cache(
+            hf_cache,
+            "unsloth/vision-GGUF",
+            {"mmproj-vision-F16.gguf": 10},
+            snapshot_sha = "b" * 40,
+        )
+        os.utime(old, (1000, 1000))
+        os.utime(new, (2000, 2000))
+
+        out = _list_gguf_variants_from_hf_cache("unsloth/vision-GGUF")
+        assert out is not None
+        variants, has_vision = out
+        assert [v.quant for v in variants] == ["Q4_K_M"]
+        assert has_vision is True
+
+    def test_companion_only_cache_returns_empty_variants_with_vision(self, hf_cache):
+        # Only a vision projector is cached anywhere: report the vision flag
+        # with an empty variant list rather than None.
+        _build_cache(hf_cache, "unsloth/vision-GGUF", {"mmproj-vision-F16.gguf": 10})
+        out = _list_gguf_variants_from_hf_cache("unsloth/vision-GGUF")
+        assert out is not None
+        variants, has_vision = out
+        assert variants == []
+        assert has_vision is True
+
+
 class TestListGgufVariantsOffline:
     def test_offline_env_short_circuits_api(self, hf_cache, clean_offline_env, monkeypatch):
         _build_cache(hf_cache, "unsloth/a", {"a-UD-Q4_K_XL.gguf": 1})
diff --git a/studio/backend/tests/test_openai_auto_switch.py b/studio/backend/tests/test_openai_auto_switch.py
index 7d2e2213b3..d02a2a4f7e 100644
--- a/studio/backend/tests/test_openai_auto_switch.py
+++ b/studio/backend/tests/test_openai_auto_switch.py
@@ -3037,3 +3037,60 @@ def test_acquire_swap_gate_is_cancellation_safe():
         inference_route._auto_switch_process_lock.release()
 
     asyncio.run(asyncio.wait_for(main(), timeout = 5))
+
+
+def test_no_model_loaded_detail_appends_hint_only_when_off(monkeypatch):
+    # The "no model loaded" errors point at the opt-in auto-switch toggle so a
+    # request naming a listed-but-unloaded model is self-explanatory -- but only
+    # when it's off. With it on the name simply didn't resolve, so no hint.
+    base = "No GGUF model loaded. Load a GGUF model first."
+
+    monkeypatch.setattr(settings, "get_openai_auto_switch_enabled", lambda: False)
+    off = inference_route._no_model_loaded_detail(base)
+    assert off.startswith(base)
+    assert "Model auto-switch" in off and "Settings > API" in off
+
+    monkeypatch.setattr(settings, "get_openai_auto_switch_enabled", lambda: True)
+    assert inference_route._no_model_loaded_detail(base) == base
+
+
+def _run_responses_stream_no_model(monkeypatch, *, enabled, active_model_name):
+    # Drive _responses_stream's GGUF-not-loaded guard: llama backend unloaded,
+    # inference backend maybe holding a non-GGUF model. Returns the 400 detail.
+    from fastapi import HTTPException
+    from models.inference import ResponsesRequest, ChatMessage
+
+    monkeypatch.setattr(settings, "get_openai_auto_switch_enabled", lambda: enabled)
+    monkeypatch.setattr(
+        inference_route, "get_llama_cpp_backend", lambda: _FakeBackend(loaded_id = None)
+    )
+    monkeypatch.setattr(
+        inference_route,
+        "get_inference_backend",
+        lambda: type("_B", (), {"active_model_name": active_model_name})(),
+    )
+    payload = ResponsesRequest(model = "unsloth/Qwen3.5-4B-GGUF", stream = True)
+    messages = [ChatMessage(role = "user", content = "hi")]
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(inference_route._responses_stream(payload, messages, None))
+    assert exc.value.status_code == 400
+    return exc.value.detail
+
+
+def test_responses_stream_hint_matches_toggle_regardless_of_active_model(monkeypatch):
+    # Streaming /v1/responses shares the GGUF-only 400 with the other "no model
+    # loaded" sites, so the auto-switch hint attaches whenever the toggle is
+    # off -- including while a non-GGUF model is active, since auto-switch
+    # evicts it to load a resolved GGUF (_maybe_auto_switch_model's resolver
+    # branch has no active-model guard, unlike its reload-stash branch). Only
+    # the toggle being on suppresses it.
+    hinted = _run_responses_stream_no_model(monkeypatch, enabled = False, active_model_name = None)
+    assert "Model auto-switch" in hinted
+
+    on = _run_responses_stream_no_model(monkeypatch, enabled = True, active_model_name = None)
+    assert "Model auto-switch" not in on
+
+    non_gguf_loaded = _run_responses_stream_no_model(
+        monkeypatch, enabled = False, active_model_name = "unsloth/Llama-3.2-1B-Instruct"
+    )
+    assert "Model auto-switch" in non_gguf_loaded
diff --git a/studio/backend/tests/test_openai_tool_passthrough.py b/studio/backend/tests/test_openai_tool_passthrough.py
index 05e017ba7f..910818d7d8 100644
--- a/studio/backend/tests/test_openai_tool_passthrough.py
+++ b/studio/backend/tests/test_openai_tool_passthrough.py
@@ -8,6 +8,7 @@ import sys
 import asyncio
 import json
 import threading
+import time
 from types import SimpleNamespace
 
 _backend = os.path.join(os.path.dirname(__file__), "..")
@@ -29,7 +30,17 @@ from core.inference.anthropic_compat import (
     anthropic_tool_choice_to_openai,
 )
 from core.inference.api_monitor import ApiMonitor
+from core.inference.llama_admission import (
+    ADMISSION_KEEPALIVE_INTERVAL_ENV,
+    ADMISSION_MAX_QUEUE_ENV,
+    ADMISSION_QUEUE_TIMEOUT_ENV,
+    LlamaAdmissionCancelled,
+    LlamaAdmissionConfig,
+    get_llama_admission_queue,
+    reset_llama_admission_queues,
+)
 from routes.inference import (
+    _aclose_stream_resources,
     _build_chat_request,
     _build_openai_passthrough_body,
     _build_passthrough_payload,
@@ -38,24 +49,69 @@ from routes.inference import (
     _coalesce_consecutive_user_turns,
     _drop_empty_assistant_sentinels,
     _effective_max_tokens,
+    _effective_openai_max_tokens,
+    _effective_openai_max_tokens_from_values,
     _extract_content_parts,
     _friendly_error,
     _friendly_upstream_error,
     _merge_user_content,
     _monitor_openai_chunk,
     _monitor_openai_sse_event,
+    _normalize_openai_passthrough_sse_line,
+    _openai_compat_stream_stall_timeout,
+    _openai_llama_admission_capacity,
     _openai_messages_for_gguf_chat,
+    _openai_passthrough_sse_line_terminal_state,
+    _openai_passthrough_upstream_headers,
     _openai_passthrough_non_streaming,
     _openai_passthrough_stream,
+    _responses_stream,
+    _openai_stream_error_sse,
     _openai_stream_usage_chunk,
+    _openai_admission_wait_stream_chunks,
+    _wait_for_openai_admission_non_streaming,
     _proxy_to_external_provider,
     _SameTaskStreamingResponse,
+    _OPENAI_COMPAT_STREAM_STALL_TIMEOUT_ENV,
     _set_or_prepend_system_message,
     openai_completions,
     openai_embeddings,
     openai_chat_completions,
 )
-from state.tool_policy import reset_tool_policy
+from state.tool_policy import reset_tool_policy, set_tool_policy
+
+
+@pytest.fixture(autouse = True)
+def _reset_admission_queues():
+    reset_llama_admission_queues()
+    yield
+    reset_llama_admission_queues()
+
+
+def test_aclose_stream_resources_attempts_remaining_closes_after_cancel():
+    class Closeable:
+        def __init__(self, *, cancel = False):
+            self.cancel = cancel
+            self.closed = False
+
+        async def aclose(self):
+            self.closed = True
+            if self.cancel:
+                raise asyncio.CancelledError()
+
+    async def _run():
+        iterator = Closeable(cancel = True)
+        resp = Closeable()
+        client = Closeable()
+
+        with pytest.raises(asyncio.CancelledError):
+            await _aclose_stream_resources(iterator = iterator, resp = resp, client = client)
+
+        assert iterator.closed
+        assert resp.closed
+        assert client.closed
+
+    asyncio.run(_run())
 
 
 class TestFriendlyUpstreamError:
@@ -548,6 +604,263 @@ class TestChatCompletionRequestToolFields:
         assert "n > 1 is not supported" in entry["error"]
         assert monitor.active_count() == 0
 
+    def test_client_tools_rejected_when_gguf_template_has_no_tool_support(self, monkeypatch):
+        import routes.inference as inference_route
+
+        class _GGUFBackend:
+            is_loaded = True
+            model_identifier = "test-gguf"
+            supports_tools = False
+            is_vision = False
+            _is_audio = False
+            context_length = 4096
+
+            def generate_chat_completion(self, **_kwargs):
+                raise AssertionError("client tools must not fall through to the standard GGUF path")
+
+        monitor = ApiMonitor(max_entries = 3)
+        monkeypatch.setattr(inference_route, "api_monitor", monitor)
+        client = self._v1_client(monkeypatch, _GGUFBackend())
+        resp = client.post(
+            "/v1/chat/completions",
+            json = {
+                "messages": [{"role": "user", "content": "hi"}],
+                "tools": [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object"},
+                        },
+                    }
+                ],
+            },
+        )
+
+        self._assert_unsupported_param(resp, "tools")
+        assert "does not advertise tools" in resp.json()["error"]["message"]
+        [entry] = monitor.snapshot()
+        assert entry["status"] == "error"
+        assert "does not advertise tools" in entry["error"]
+        assert monitor.active_count() == 0
+
+    def test_client_tools_use_passthrough_capability_when_tool_loop_is_disabled(self, monkeypatch):
+        import routes.inference as inference_route
+
+        captured = {}
+
+        class _GGUFBackend:
+            is_loaded = True
+            model_identifier = "test-gguf"
+            supports_tools = False
+            supports_tool_passthrough = True
+            is_vision = False
+            _is_audio = False
+            context_length = 4096
+            base_url = "http://llama.passthrough-capability.test"
+            _request_reasoning_kwargs = lambda *_args, **_kwargs: None
+
+            def generate_chat_completion(self, **_kwargs):
+                raise AssertionError("client tools must use passthrough")
+
+            def generate_chat_completion_with_tools(self, **_kwargs):
+                raise AssertionError("Studio tool loop must stay disabled")
+
+        async def fake_passthrough(llama_backend, payload, model_name, **kwargs):
+            captured["body"] = inference_route._build_openai_passthrough_body(
+                payload,
+                backend_ctx = llama_backend.context_length,
+                llama_backend = llama_backend,
+            )
+            inference_route.api_monitor.finish(kwargs.get("monitor_id"))
+            return inference_route.JSONResponse({"ok": True, "model": model_name})
+
+        monitor = ApiMonitor(max_entries = 3)
+        monkeypatch.setattr(inference_route, "api_monitor", monitor)
+        monkeypatch.setattr(
+            inference_route,
+            "_openai_passthrough_non_streaming",
+            fake_passthrough,
+        )
+        client = self._v1_client(monkeypatch, _GGUFBackend())
+        resp = client.post(
+            "/v1/chat/completions",
+            json = {
+                "messages": [{"role": "user", "content": "use client tool"}],
+                "tools": [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object"},
+                        },
+                    }
+                ],
+            },
+        )
+
+        assert resp.status_code == 200
+        assert resp.json()["ok"] is True
+        assert captured["body"]["tools"][0]["function"]["name"] == "lookup"
+        [entry] = monitor.snapshot()
+        assert entry["status"] == "completed"
+        assert monitor.active_count() == 0
+
+    def test_enable_tools_on_non_tool_backend_keeps_client_tools_on_passthrough(self, monkeypatch):
+        # DiffusionGemma forces supports_tools off while passthrough stays
+        # available (#6851): enable_tools=True must not steal client tools
+        # from the passthrough into a Studio tool loop that cannot run.
+        import routes.inference as inference_route
+
+        captured = {}
+
+        class _GGUFBackend:
+            is_loaded = True
+            model_identifier = "test-gguf"
+            supports_tools = False
+            supports_tool_passthrough = True
+            is_vision = False
+            _is_audio = False
+            context_length = 4096
+            base_url = "http://llama.passthrough-capability.test"
+            _request_reasoning_kwargs = lambda *_args, **_kwargs: None
+
+            def generate_chat_completion(self, **_kwargs):
+                raise AssertionError("client tools must use passthrough")
+
+            def generate_chat_completion_with_tools(self, **_kwargs):
+                raise AssertionError("Studio tool loop cannot run on a non-tool backend")
+
+        async def fake_passthrough(llama_backend, payload, model_name, **kwargs):
+            captured["body"] = inference_route._build_openai_passthrough_body(
+                payload,
+                backend_ctx = llama_backend.context_length,
+                llama_backend = llama_backend,
+            )
+            inference_route.api_monitor.finish(kwargs.get("monitor_id"))
+            return inference_route.JSONResponse({"ok": True, "model": model_name})
+
+        monitor = ApiMonitor(max_entries = 3)
+        monkeypatch.setattr(inference_route, "api_monitor", monitor)
+        monkeypatch.setattr(
+            inference_route,
+            "_openai_passthrough_non_streaming",
+            fake_passthrough,
+        )
+        client = self._v1_client(monkeypatch, _GGUFBackend())
+        resp = client.post(
+            "/v1/chat/completions",
+            json = {
+                "messages": [{"role": "user", "content": "use client tool"}],
+                "enable_tools": True,
+                "tools": [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object"},
+                        },
+                    }
+                ],
+            },
+        )
+
+        assert resp.status_code == 200
+        assert resp.json()["ok"] is True
+        assert captured["body"]["tools"][0]["function"]["name"] == "lookup"
+        [entry] = monitor.snapshot()
+        assert entry["status"] == "completed"
+        assert monitor.active_count() == 0
+
+    def test_tool_choice_none_allows_tool_catalog_without_tool_template(self, monkeypatch):
+        import routes.inference as inference_route
+
+        class _GGUFBackend:
+            is_loaded = True
+            model_identifier = "test-gguf"
+            supports_tools = False
+            is_vision = False
+            _is_audio = False
+            context_length = 4096
+
+            def generate_chat_completion(self, **kwargs):
+                assert kwargs["max_tokens"] is None
+                yield "plain response"
+
+        monitor = ApiMonitor(max_entries = 3)
+        monkeypatch.setattr(inference_route, "api_monitor", monitor)
+        client = self._v1_client(monkeypatch, _GGUFBackend())
+        resp = client.post(
+            "/v1/chat/completions",
+            json = {
+                "messages": [{"role": "user", "content": "hi"}],
+                "tools": [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object"},
+                        },
+                    }
+                ],
+                "tool_choice": "none",
+            },
+        )
+
+        assert resp.status_code == 200
+        assert resp.json()["choices"][0]["message"]["content"] == "plain response"
+        [entry] = monitor.snapshot()
+        assert entry["status"] == "completed"
+        assert entry["reply"] == "plain response"
+        assert monitor.active_count() == 0
+
+    def test_tool_call_history_rejected_when_gguf_template_has_no_tool_support(self, monkeypatch):
+        import routes.inference as inference_route
+
+        class _GGUFBackend:
+            is_loaded = True
+            model_identifier = "test-gguf"
+            supports_tools = False
+            is_vision = False
+            _is_audio = False
+            context_length = 4096
+
+            def generate_chat_completion(self, **_kwargs):
+                raise AssertionError(
+                    "tool-call history must not fall through to the standard GGUF path"
+                )
+
+        monitor = ApiMonitor(max_entries = 3)
+        monkeypatch.setattr(inference_route, "api_monitor", monitor)
+        client = self._v1_client(monkeypatch, _GGUFBackend())
+        resp = client.post(
+            "/v1/chat/completions",
+            json = {
+                "messages": [
+                    {"role": "user", "content": "use a tool"},
+                    {
+                        "role": "assistant",
+                        "content": None,
+                        "tool_calls": [
+                            {
+                                "id": "call_1",
+                                "type": "function",
+                                "function": {"name": "lookup", "arguments": "{}"},
+                            }
+                        ],
+                    },
+                    {"role": "tool", "tool_call_id": "call_1", "content": "{}"},
+                ],
+            },
+        )
+
+        self._assert_unsupported_param(resp, "messages")
+        assert "does not advertise tools" in resp.json()["error"]["message"]
+        [entry] = monitor.snapshot()
+        assert entry["status"] == "error"
+        assert "does not advertise tools" in entry["error"]
+        assert monitor.active_count() == 0
+
     def test_n_rejected_for_non_gguf_path(self, monkeypatch):
         class _NoGGUFBackend:
             is_loaded = False
@@ -748,11 +1061,32 @@ class TestBuildPassthroughPayloadToolChoice:
         )
         assert body.get("stream_options") == {"include_usage": False}
 
+    def test_response_format_without_tools_omits_tool_fields(self):
+        args = self._args()
+        args["openai_tools"] = None
+
+        body = _build_passthrough_payload(
+            **args,
+            response_format = {"type": "json_object"},
+        )
+
+        assert body["response_format"] == {"type": "json_object"}
+        assert "tools" not in body
+        assert "tool_choice" not in body
+
     def test_repetition_penalty_renamed(self):
         body = _build_passthrough_payload(**self._args(), repetition_penalty = 1.1)
         assert body.get("repeat_penalty") == 1.1
         assert "repetition_penalty" not in body
 
+    def test_omitted_passthrough_max_tokens_uses_backend_context(self):
+        args = self._args()
+        args["max_tokens"] = None
+
+        body = _build_passthrough_payload(**args, backend_ctx = 4096)
+
+        assert body["max_tokens"] == 4096
+
     def test_passthrough_body_merges_system_and_developer_messages(self):
         payload = ChatCompletionRequest(
             model = "default",
@@ -772,6 +1106,74 @@ class TestBuildPassthroughPayloadToolChoice:
         ]
 
 
+class TestOpenAIPassthroughSSETerminalState:
+    def test_done_sentinel(self):
+        assert _openai_passthrough_sse_line_terminal_state("data: [DONE]") == "done"
+
+    def test_finish_reason_with_space(self):
+        line = 'data: {"choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}'
+        assert _openai_passthrough_sse_line_terminal_state(line) == "finish"
+
+    def test_finish_reason_without_space(self):
+        line = 'data:{"choices":[{"index":0,"delta":{},"finish_reason":"tool_calls"}]}'
+        assert _openai_passthrough_sse_line_terminal_state(line) == "finish"
+
+    def test_usage_chunk(self):
+        line = 'data: {"choices":[],"usage":{"prompt_tokens":1,"completion_tokens":2}}'
+        assert _openai_passthrough_sse_line_terminal_state(line) == "usage"
+
+    def test_error_chunk(self):
+        line = 'data: {"error":{"message":"boom"}}'
+        assert _openai_passthrough_sse_line_terminal_state(line) == "error"
+
+    def test_cap_parallel_tool_calls_accepts_no_space_after_data_colon(self):
+        line = (
+            'data:{"choices":[{"delta":{"tool_calls":['
+            '{"index":0,"function":{"name":"a"}},'
+            '{"index":1,"function":{"name":"b"}}]}}]}'
+        )
+
+        capped = _normalize_openai_passthrough_sse_line(line, cap_parallel_tool_calls = True)
+
+        data = json.loads(capped[len("data:") :].lstrip())
+        assert data["choices"][0]["delta"]["tool_calls"] == [
+            {"index": 0, "function": {"name": "a"}}
+        ]
+
+    def test_plain_content_line_is_returned_identically(self):
+        # The relay dispatches terminal classification on `out_line is raw_line`,
+        # so the no-mutation path must return the identical string object.
+        line = 'data: {"choices":[{"index":0,"delta":{"content":"hello"},"finish_reason":null}]}'
+        assert _normalize_openai_passthrough_sse_line(line) is line
+        assert _normalize_openai_passthrough_sse_line(line, cap_parallel_tool_calls = True) is line
+
+    def test_reasoning_key_inside_content_text_keeps_line_identical(self):
+        # Fast-path substring gate fires, but the parse finds nothing to change:
+        # the original object must come back so the relay stays byte-identical.
+        line = (
+            'data: {"choices":[{"index":0,"delta":{"content":'
+            '"mentions \\"reasoning_content\\" in text"},"finish_reason":null}]}'
+        )
+        assert _normalize_openai_passthrough_sse_line(line) is line
+
+    def test_reasoning_only_delta_gets_empty_content(self):
+        line = (
+            'data: {"choices":[{"index":0,'
+            '"delta":{"reasoning_content":"thinking"},'
+            '"finish_reason":null}]}'
+        )
+
+        normalized = _normalize_openai_passthrough_sse_line(line)
+
+        data = json.loads(normalized[len("data:") :].lstrip())
+        delta = data["choices"][0]["delta"]
+        assert delta["reasoning_content"] == "thinking"
+        assert delta["content"] == ""
+
+    def test_reasoning_normalization_preserves_done_sentinel(self):
+        assert _normalize_openai_passthrough_sse_line("data: [DONE]") == "data: [DONE]"
+
+
 # =====================================================================
 # Passthrough reasoning kwargs — enable_thinking / reasoning_effort /
 # preserve_thinking must reach llama-server via chat_template_kwargs,
@@ -892,6 +1294,172 @@ class TestOpenAICompatibilityHelpers:
         payload = SimpleNamespace(max_tokens = 128, max_completion_tokens = 64)
         assert _effective_max_tokens(payload) == 64
 
+    def test_openai_compat_max_tokens_returns_none_when_omitted(self):
+        payload = SimpleNamespace(max_tokens = None, max_completion_tokens = None)
+        assert _effective_openai_max_tokens(payload) is None
+
+    @pytest.mark.parametrize(
+        ("payload", "expected"),
+        [
+            (SimpleNamespace(max_tokens = 8192, max_completion_tokens = None), 8192),
+            (SimpleNamespace(max_tokens = 8192, max_completion_tokens = 256), 256),
+        ],
+    )
+    def test_openai_compat_explicit_values_pass_through(self, payload, expected):
+        assert _effective_openai_max_tokens(payload) == expected
+
+    @pytest.mark.parametrize(
+        ("payload", "param"),
+        [
+            (SimpleNamespace(max_tokens = "128", max_completion_tokens = None), "max_tokens"),
+            (SimpleNamespace(max_tokens = True, max_completion_tokens = None), "max_tokens"),
+            (SimpleNamespace(max_tokens = 12.5, max_completion_tokens = None), "max_tokens"),
+            (
+                SimpleNamespace(max_tokens = None, max_completion_tokens = "128"),
+                "max_completion_tokens",
+            ),
+        ],
+    )
+    def test_openai_compat_max_tokens_rejects_non_integer_explicit_values(self, payload, param):
+        with pytest.raises(HTTPException) as exc:
+            _effective_openai_max_tokens(payload)
+
+        assert exc.value.status_code == 400
+        assert exc.value.detail["error"]["param"] == param
+        assert exc.value.detail["error"]["code"] == "invalid_type"
+
+    def test_openai_compat_max_tokens_zero_is_valid_and_negative_rejected(self):
+        # Legacy completions spec: max_tokens has minimum 0, so 0 must pass
+        # through; only negatives are invalid_value.
+        assert _effective_openai_max_tokens_from_values(0) == 0
+
+        with pytest.raises(HTTPException) as exc:
+            _effective_openai_max_tokens_from_values(-1)
+
+        assert exc.value.status_code == 400
+        assert exc.value.detail["error"]["code"] == "invalid_value"
+        assert exc.value.detail["error"]["param"] == "max_tokens"
+
+    def test_chat_reasoning_chunk_carries_empty_content(self):
+        from routes.inference import _chat_reasoning_chunk
+
+        line = _chat_reasoning_chunk("chatcmpl-test", 123, "gguf", "thinking...")
+        chunk = json.loads(line[len("data: ") :])
+        delta = chunk["choices"][0]["delta"]
+
+        assert delta["reasoning_content"] == "thinking..."
+        assert delta["content"] == ""
+
+    def test_passthrough_upstream_headers_include_backend_auth(self):
+        headers = _openai_passthrough_upstream_headers(
+            llama_backend = SimpleNamespace(_auth_headers = {"Authorization": "Bearer secret"}),
+        )
+
+        assert headers["Authorization"] == "Bearer secret"
+        assert headers["Connection"] == "close"
+
+    def test_openai_admission_capacity_prefers_backend_effective_slots(self):
+        request = SimpleNamespace(
+            app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+        )
+        backend = SimpleNamespace(effective_parallel_slots = 3)
+
+        assert _openai_llama_admission_capacity(request, backend) == 3
+
+    @pytest.mark.parametrize("backend_value", [None, 0, -1, "not-an-int"])
+    def test_openai_admission_capacity_falls_back_to_app_state(self, backend_value):
+        request = SimpleNamespace(
+            app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 2))
+        )
+        backend = SimpleNamespace(effective_parallel_slots = backend_value)
+
+        assert _openai_llama_admission_capacity(request, backend) == 2
+
+    def test_openai_admission_capacity_falls_back_to_one_without_request(self):
+        assert _openai_llama_admission_capacity(None, SimpleNamespace()) == 1
+
+    def test_openai_admission_non_streaming_exits_invalidated_waiter(self):
+        async def _run():
+            queue = get_llama_admission_queue("http://llama.invalidated.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+            reservation = queue.reserve(capacity = 1, config = LlamaAdmissionConfig())
+            assert reservation._waiter is not None
+
+            reservation._waiter.future.cancel()
+
+            with pytest.raises(LlamaAdmissionCancelled):
+                await asyncio.wait_for(
+                    _wait_for_openai_admission_non_streaming(
+                        reservation,
+                        LlamaAdmissionConfig(),
+                        request = None,
+                        cancel_event = None,
+                    ),
+                    timeout = 0.1,
+                )
+
+            blocker.release()
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+
+        asyncio.run(_run())
+
+    def test_openai_admission_stream_exits_invalidated_waiter(self):
+        async def _run():
+            queue = get_llama_admission_queue("http://llama.invalidated.stream.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+            reservation = queue.reserve(capacity = 1, config = LlamaAdmissionConfig())
+            assert reservation._waiter is not None
+
+            reservation._waiter.future.cancel()
+
+            chunks = _openai_admission_wait_stream_chunks(
+                reservation,
+                LlamaAdmissionConfig(),
+                request = None,
+                cancel_event = None,
+            )
+            with pytest.raises(LlamaAdmissionCancelled):
+                await asyncio.wait_for(chunks.__anext__(), timeout = 0.1)
+
+            blocker.release()
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+
+        asyncio.run(_run())
+
+    def test_openai_compat_stream_stall_timeout_uses_default(self, monkeypatch):
+        monkeypatch.delenv(_OPENAI_COMPAT_STREAM_STALL_TIMEOUT_ENV, raising = False)
+        assert _openai_compat_stream_stall_timeout() == 120.0
+
+    def test_openai_compat_stream_stall_timeout_uses_env_override(self, monkeypatch):
+        monkeypatch.setenv(_OPENAI_COMPAT_STREAM_STALL_TIMEOUT_ENV, "4.5")
+        assert _openai_compat_stream_stall_timeout() == 4.5
+
+    @pytest.mark.parametrize("raw_value", ["", "not-a-float"])
+    def test_openai_compat_stream_stall_timeout_invalid_env_uses_default(
+        self, monkeypatch, raw_value
+    ):
+        monkeypatch.setenv(_OPENAI_COMPAT_STREAM_STALL_TIMEOUT_ENV, raw_value)
+        assert _openai_compat_stream_stall_timeout() == 120.0
+
+    @pytest.mark.parametrize("raw_value", ["0", "-1"])
+    def test_openai_compat_stream_stall_timeout_non_positive_env_disables(
+        self, monkeypatch, raw_value
+    ):
+        monkeypatch.setenv(_OPENAI_COMPAT_STREAM_STALL_TIMEOUT_ENV, raw_value)
+        assert _openai_compat_stream_stall_timeout() is None
+
+    def test_openai_stream_error_sse_closes_with_done(self):
+        error = {"error": {"message": "boom"}}
+        assert _openai_stream_error_sse(error) == (
+            'data: {"error": {"message": "boom"}}\n\n' "data: [DONE]\n\n"
+        )
+
     @pytest.mark.parametrize(
         "finish_reason",
         ["stop", "length", "tool_calls", "content_filter", "function_call"],
@@ -1515,9 +2083,659 @@ class TestGgufVisionToolRouting:
         assert "".join(d.get("reasoning_content", "") for d in deltas) == "plan"
         assert "".join(d.get("content", "") for d in deltas) == "visible"
         assert all("<think>" not in d.get("content", "") for d in deltas)
+        assert all("content" in d for d in deltas if "reasoning_content" in d)
         [entry] = result.monitor.snapshot()
         assert entry["reply"] == "visible"
 
+    def test_standard_gguf_stream_queued_request_sends_keepalive_before_generation(
+        self, monkeypatch
+    ):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request(self._Request):
+                app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+
+            def _generate(**_kwargs):
+                raise AssertionError("standard GGUF generation must not start while queued")
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = False,
+                supports_reasoning = True,
+                reasoning_always_on = True,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                base_url = "http://llama.standard.test",
+                effective_parallel_slots = 1,
+                generate_chat_completion = _generate,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setenv(ADMISSION_KEEPALIVE_INTERVAL_ENV, "0.01")
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+
+            queue = get_llama_admission_queue("http://llama.standard.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+                stream = True,
+            )
+            response = await openai_chat_completions(
+                payload,
+                request = Request(),
+                current_subject = "test",
+            )
+            iterator = response.body_iterator
+            try:
+                chunk = await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+                assert chunk == ": keep-alive\n\n"
+                snapshot = queue.snapshot()
+                assert snapshot.active == 1
+                assert snapshot.queued == 1
+            finally:
+                aclose = getattr(iterator, "aclose", None)
+                if aclose is not None:
+                    await aclose()
+                blocker.release()
+
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_standard_gguf_stream_close_after_first_chunk_cleans_tracker(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            cancel_id = "standard-stream-close-cleanup"
+
+            def _generate(**_kwargs):
+                yield "visible"
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = False,
+                supports_reasoning = True,
+                reasoning_always_on = True,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                base_url = "http://llama.standard.test",
+                effective_parallel_slots = 1,
+                generate_chat_completion = _generate,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+                stream = True,
+                cancel_id = cancel_id,
+            )
+            response = await openai_chat_completions(
+                payload,
+                request = self._Request(),
+                current_subject = "test",
+            )
+            iterator = response.body_iterator
+            assert cancel_id in inf_mod._CANCEL_REGISTRY
+            await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+            aclose = getattr(iterator, "aclose", None)
+            assert aclose is not None
+            await aclose()
+
+            assert cancel_id not in inf_mod._CANCEL_REGISTRY
+            assert get_llama_admission_queue("http://llama.standard.test").snapshot().active == 0
+
+        asyncio.run(_run())
+
+    def test_standard_gguf_stream_task_cancel_after_first_chunk_finalizes_monitor(
+        self, monkeypatch
+    ):
+        async def _run():
+            import routes.inference as inf_mod
+
+            started = threading.Event()
+            released = threading.Event()
+
+            def _generate(**kwargs):
+                cancel_event = kwargs["cancel_event"]
+                started.set()
+                while not cancel_event.is_set():
+                    time.sleep(0.005)
+                released.set()
+                yield from ()
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = False,
+                supports_reasoning = True,
+                reasoning_always_on = True,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                base_url = "http://llama.standard.test",
+                effective_parallel_slots = 1,
+                generate_chat_completion = _generate,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+                stream = True,
+            )
+            response = await openai_chat_completions(
+                payload,
+                request = self._Request(),
+                current_subject = "test",
+            )
+            iterator = response.body_iterator
+            assert await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+            pending = asyncio.create_task(iterator.__anext__())
+            assert await asyncio.to_thread(started.wait, 1.0)
+
+            await asyncio.sleep(0)
+            pending.cancel()
+            with pytest.raises(asyncio.CancelledError):
+                await asyncio.wait_for(pending, timeout = 1.0)
+
+            assert released.is_set()
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+            assert get_llama_admission_queue("http://llama.standard.test").snapshot().active == 0
+
+        asyncio.run(_run())
+
+    def test_gguf_tool_stream_queued_request_sends_keepalive_before_generation(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request(self._Request):
+                app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+
+            async def fake_select_tools(*_args, **_kwargs):
+                return [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ]
+
+            def _generate(**_kwargs):
+                raise AssertionError("GGUF tool loop must not start while queued")
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = True,
+                supports_reasoning = True,
+                reasoning_always_on = True,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                base_url = "http://llama.tool.test",
+                effective_parallel_slots = 1,
+                generate_chat_completion = lambda **_kwargs: "unused",
+                generate_chat_completion_with_tools = _generate,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setenv(ADMISSION_KEEPALIVE_INTERVAL_ENV, "0.01")
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+            monkeypatch.setattr(inf_mod, "_select_request_tools", fake_select_tools)
+
+            queue = get_llama_admission_queue("http://llama.tool.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+                enable_tools = True,
+                stream = True,
+            )
+            response = await openai_chat_completions(
+                payload,
+                request = Request(),
+                current_subject = "test",
+            )
+            iterator = response.body_iterator
+            try:
+                chunk = await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+                assert chunk == ": keep-alive\n\n"
+                snapshot = queue.snapshot()
+                assert snapshot.active == 1
+                assert snapshot.queued == 1
+            finally:
+                aclose = getattr(iterator, "aclose", None)
+                if aclose is not None:
+                    await aclose()
+                blocker.release()
+
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_gguf_tool_stream_task_cancel_after_first_chunk_finalizes_monitor(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            async def fake_select_tools(*_args, **_kwargs):
+                return [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ]
+
+            started = threading.Event()
+            released = threading.Event()
+
+            def _tools(**kwargs):
+                cancel_event = kwargs["cancel_event"]
+                started.set()
+                while not cancel_event.is_set():
+                    time.sleep(0.005)
+                released.set()
+                yield from ()
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = True,
+                supports_reasoning = True,
+                reasoning_always_on = True,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                base_url = "http://llama.tool.test",
+                effective_parallel_slots = 1,
+                generate_chat_completion = lambda **_kwargs: "unused",
+                generate_chat_completion_with_tools = _tools,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+            monkeypatch.setattr(inf_mod, "_select_request_tools", fake_select_tools)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+                enable_tools = True,
+                stream = True,
+            )
+            response = await openai_chat_completions(
+                payload,
+                request = self._Request(),
+                current_subject = "test",
+            )
+            iterator = response.body_iterator
+            assert await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+            pending = asyncio.create_task(iterator.__anext__())
+            assert await asyncio.to_thread(started.wait, 1.0)
+
+            await asyncio.sleep(0)
+            pending.cancel()
+            with pytest.raises(asyncio.CancelledError):
+                await asyncio.wait_for(pending, timeout = 1.0)
+
+            assert released.is_set()
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+            assert get_llama_admission_queue("http://llama.tool.test").snapshot().active == 0
+
+        asyncio.run(_run())
+
+    def test_global_enable_tools_does_not_preempt_response_format_passthrough(self, monkeypatch):
+        import routes.inference as inf_mod
+
+        reset_tool_policy()
+        set_tool_policy(True)
+        captured = {}
+
+        def _plain(**_kwargs):
+            raise AssertionError("plain GGUF path should not be used")
+
+        def _tools(**_kwargs):
+            raise AssertionError("Studio tool loop should not steal response_format")
+
+        backend = SimpleNamespace(
+            is_loaded = True,
+            is_vision = False,
+            supports_tools = True,
+            _is_audio = False,
+            model_identifier = "test-gguf",
+            context_length = 4096,
+            base_url = "http://llama.policy.test",
+            _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+            generate_chat_completion = _plain,
+            generate_chat_completion_with_tools = _tools,
+        )
+
+        async def fake_passthrough(llama_backend, payload, model_name, **_kwargs):
+            captured["body"] = inf_mod._build_openai_passthrough_body(
+                payload,
+                backend_ctx = llama_backend.context_length,
+                llama_backend = llama_backend,
+            )
+            return inf_mod.JSONResponse({"ok": True, "model": model_name})
+
+        try:
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+            monkeypatch.setattr(
+                inf_mod,
+                "_openai_passthrough_non_streaming",
+                fake_passthrough,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "json"}],
+                response_format = {"type": "json_object"},
+            )
+            response = self._drive(
+                openai_chat_completions(
+                    payload,
+                    request = self._Request(),
+                    current_subject = "test",
+                )
+            )
+
+            assert json.loads(response.body)["ok"] is True
+            assert captured["body"]["response_format"] == {"type": "json_object"}
+            assert "tools" not in captured["body"]
+            assert "tool_choice" not in captured["body"]
+        finally:
+            reset_tool_policy()
+
+    def test_global_enable_tools_does_not_replace_client_tools_passthrough(self, monkeypatch):
+        import routes.inference as inf_mod
+
+        reset_tool_policy()
+        set_tool_policy(True)
+        captured = {}
+        client_tools = [
+            {
+                "type": "function",
+                "function": {
+                    "name": "client_lookup",
+                    "parameters": {"type": "object", "properties": {}},
+                },
+            }
+        ]
+
+        def _plain(**_kwargs):
+            raise AssertionError("plain GGUF path should not be used")
+
+        def _tools(**_kwargs):
+            raise AssertionError("Studio tool loop should not replace client tools")
+
+        backend = SimpleNamespace(
+            is_loaded = True,
+            is_vision = False,
+            supports_tools = True,
+            _is_audio = False,
+            model_identifier = "test-gguf",
+            context_length = 4096,
+            base_url = "http://llama.policy.test",
+            _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+            generate_chat_completion = _plain,
+            generate_chat_completion_with_tools = _tools,
+        )
+
+        async def fake_passthrough(llama_backend, payload, model_name, **_kwargs):
+            captured["body"] = inf_mod._build_openai_passthrough_body(
+                payload,
+                backend_ctx = llama_backend.context_length,
+                llama_backend = llama_backend,
+            )
+            return inf_mod.JSONResponse({"ok": True, "model": model_name})
+
+        try:
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+            monkeypatch.setattr(
+                inf_mod,
+                "_openai_passthrough_non_streaming",
+                fake_passthrough,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "use client tool"}],
+                tools = client_tools,
+            )
+            response = self._drive(
+                openai_chat_completions(
+                    payload,
+                    request = self._Request(),
+                    current_subject = "test",
+                )
+            )
+
+            assert json.loads(response.body)["ok"] is True
+            assert captured["body"]["tools"] == client_tools
+            assert captured["body"]["tool_choice"] == "auto"
+        finally:
+            reset_tool_policy()
+
+    def test_global_enable_tools_honors_client_tool_choice_none(self, monkeypatch):
+        import routes.inference as inf_mod
+
+        reset_tool_policy()
+        set_tool_policy(True)
+        client_tools = [
+            {
+                "type": "function",
+                "function": {
+                    "name": "client_lookup",
+                    "parameters": {"type": "object", "properties": {}},
+                },
+            }
+        ]
+
+        def _plain(**kwargs):
+            assert kwargs["max_tokens"] is None
+            yield "plain response"
+
+        def _tools(**_kwargs):
+            raise AssertionError("tool_choice='none' must not start Studio's tool loop")
+
+        backend = SimpleNamespace(
+            is_loaded = True,
+            is_vision = False,
+            supports_tools = True,
+            _is_audio = False,
+            model_identifier = "test-gguf",
+            context_length = 4096,
+            base_url = "http://llama.policy.test",
+            _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+            generate_chat_completion = _plain,
+            generate_chat_completion_with_tools = _tools,
+        )
+
+        try:
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "do not use tools"}],
+                tools = client_tools,
+                tool_choice = "none",
+            )
+            response = self._drive(
+                openai_chat_completions(
+                    payload,
+                    request = self._Request(),
+                    current_subject = "test",
+                )
+            )
+
+            assert json.loads(response.body)["choices"][0]["message"]["content"] == "plain response"
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "completed"
+            assert entry["reply"] == "plain response"
+            assert monitor.active_count() == 0
+        finally:
+            reset_tool_policy()
+
+    def test_enabled_tools_without_enable_tools_keeps_response_format_passthrough(
+        self, monkeypatch
+    ):
+        import routes.inference as inf_mod
+
+        reset_tool_policy()
+        captured = {}
+
+        def _plain(**_kwargs):
+            raise AssertionError("plain GGUF path should not be used")
+
+        def _tools(**_kwargs):
+            raise AssertionError("enabled_tools alone must not start Studio's tool loop")
+
+        backend = SimpleNamespace(
+            is_loaded = True,
+            is_vision = False,
+            supports_tools = True,
+            _is_audio = False,
+            model_identifier = "test-gguf",
+            context_length = 4096,
+            base_url = "http://llama.enabled-tools.test",
+            _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+            generate_chat_completion = _plain,
+            generate_chat_completion_with_tools = _tools,
+        )
+
+        async def fake_passthrough(llama_backend, payload, model_name, **_kwargs):
+            captured["body"] = inf_mod._build_openai_passthrough_body(
+                payload,
+                backend_ctx = llama_backend.context_length,
+                llama_backend = llama_backend,
+            )
+            return inf_mod.JSONResponse({"ok": True, "model": model_name})
+
+        monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+        monkeypatch.setattr(inf_mod, "_openai_passthrough_non_streaming", fake_passthrough)
+        monitor = ApiMonitor(max_entries = 3)
+        monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+
+        payload = ChatCompletionRequest(
+            model = "default",
+            messages = [{"role": "user", "content": "json"}],
+            enabled_tools = ["web_search"],
+            response_format = {"type": "json_object"},
+        )
+        response = self._drive(
+            openai_chat_completions(
+                payload,
+                request = self._Request(),
+                current_subject = "test",
+            )
+        )
+
+        assert json.loads(response.body)["ok"] is True
+        assert captured["body"]["response_format"] == {"type": "json_object"}
+
+    def test_enabled_tools_without_enable_tools_keeps_client_tools_passthrough(self, monkeypatch):
+        import routes.inference as inf_mod
+
+        reset_tool_policy()
+        captured = {}
+        client_tools = [
+            {
+                "type": "function",
+                "function": {
+                    "name": "client_lookup",
+                    "parameters": {"type": "object", "properties": {}},
+                },
+            }
+        ]
+
+        def _plain(**_kwargs):
+            raise AssertionError("plain GGUF path should not be used")
+
+        def _tools(**_kwargs):
+            raise AssertionError("enabled_tools alone must not start Studio's tool loop")
+
+        backend = SimpleNamespace(
+            is_loaded = True,
+            is_vision = False,
+            supports_tools = True,
+            _is_audio = False,
+            model_identifier = "test-gguf",
+            context_length = 4096,
+            base_url = "http://llama.enabled-tools.test",
+            _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+            generate_chat_completion = _plain,
+            generate_chat_completion_with_tools = _tools,
+        )
+
+        async def fake_passthrough(llama_backend, payload, model_name, **_kwargs):
+            captured["body"] = inf_mod._build_openai_passthrough_body(
+                payload,
+                backend_ctx = llama_backend.context_length,
+                llama_backend = llama_backend,
+            )
+            return inf_mod.JSONResponse({"ok": True, "model": model_name})
+
+        monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+        monkeypatch.setattr(inf_mod, "_openai_passthrough_non_streaming", fake_passthrough)
+        monitor = ApiMonitor(max_entries = 3)
+        monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+
+        payload = ChatCompletionRequest(
+            model = "default",
+            messages = [{"role": "user", "content": "use client tool"}],
+            enabled_tools = ["web_search"],
+            tools = client_tools,
+        )
+        response = self._drive(
+            openai_chat_completions(
+                payload,
+                request = self._Request(),
+                current_subject = "test",
+            )
+        )
+
+        assert json.loads(response.body)["ok"] is True
+        assert captured["body"]["tools"] == client_tools
+        assert captured["body"]["tool_choice"] == "auto"
+
     def test_reasoning_capable_gguf_stream_splits_reasoning_by_default(self, monkeypatch):
         def _generate(**_kwargs):
             yield "<think>plan</think>visible"
@@ -1639,6 +2857,313 @@ class TestGgufVisionToolRouting:
         [entry] = result.monitor.snapshot()
         assert entry["reply"] == "visible"
 
+    def test_standard_gguf_non_streaming_admission_timeout_before_generation(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request(self._Request):
+                app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+
+            def _generate(**_kwargs):
+                raise AssertionError("standard GGUF generation must not start while queued")
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = False,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                base_url = "http://llama.standard.test",
+                effective_parallel_slots = 1,
+                generate_chat_completion = _generate,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setenv(ADMISSION_QUEUE_TIMEOUT_ENV, "0.01")
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+
+            queue = get_llama_admission_queue("http://llama.standard.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+            )
+            try:
+                with pytest.raises(HTTPException) as exc:
+                    await openai_chat_completions(
+                        payload,
+                        request = Request(),
+                        current_subject = "test",
+                    )
+                assert exc.value.status_code == 503
+            finally:
+                blocker.release()
+
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+
+        asyncio.run(_run())
+
+    def test_standard_gguf_non_streaming_cancel_id_stops_queued_request_before_generation(
+        self, monkeypatch
+    ):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request(self._Request):
+                app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+
+            def _generate(**_kwargs):
+                raise AssertionError("standard GGUF generation must not start after cancel_id")
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = False,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                base_url = "http://llama.standard.test",
+                effective_parallel_slots = 1,
+                generate_chat_completion = _generate,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+
+            queue = get_llama_admission_queue("http://llama.standard.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+
+            cancel_id = "standard-nonstream-admission-cancel"
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+                cancel_id = cancel_id,
+            )
+            task = asyncio.create_task(
+                openai_chat_completions(
+                    payload,
+                    request = Request(),
+                    current_subject = "test",
+                )
+            )
+            try:
+                for _ in range(50):
+                    if cancel_id in inf_mod._CANCEL_REGISTRY:
+                        break
+                    await asyncio.sleep(0.01)
+                assert cancel_id in inf_mod._CANCEL_REGISTRY
+                assert inf_mod._cancel_by_cancel_id_or_stash(cancel_id) == 1
+                with pytest.raises(HTTPException) as exc:
+                    await asyncio.wait_for(task, timeout = 0.5)
+                assert exc.value.status_code == 499
+            finally:
+                if not task.done():
+                    task.cancel()
+                    with pytest.raises(asyncio.CancelledError):
+                        await task
+                blocker.release()
+
+            assert cancel_id not in inf_mod._CANCEL_REGISTRY
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_standard_gguf_non_streaming_admission_task_cancel_cleans_tracker_and_slot(
+        self, monkeypatch
+    ):
+        async def _run():
+            import routes.inference as inf_mod
+
+            cancel_id = "standard-nonstream-task-cancel"
+
+            async def fake_wait(*_args, **_kwargs):
+                raise asyncio.CancelledError()
+
+            def _generate(**_kwargs):
+                raise AssertionError("standard GGUF generation must not start after task cancel")
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = False,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                base_url = "http://llama.standard.test",
+                effective_parallel_slots = 1,
+                generate_chat_completion = _generate,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+            monkeypatch.setattr(
+                inf_mod,
+                "_wait_for_openai_admission_non_streaming",
+                fake_wait,
+            )
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+                cancel_id = cancel_id,
+            )
+            with pytest.raises(asyncio.CancelledError):
+                await openai_chat_completions(
+                    payload,
+                    request = self._Request(),
+                    current_subject = "test",
+                )
+
+            assert cancel_id not in inf_mod._CANCEL_REGISTRY
+            assert get_llama_admission_queue("http://llama.standard.test").snapshot().active == 0
+
+        asyncio.run(_run())
+
+    def test_gguf_tool_non_streaming_admission_timeout_before_generation(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request(self._Request):
+                app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+
+            async def fake_select_tools(*_args, **_kwargs):
+                return [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ]
+
+            def _generate(**_kwargs):
+                raise AssertionError("GGUF tool loop must not start while queued")
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = True,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                base_url = "http://llama.tool.test",
+                effective_parallel_slots = 1,
+                generate_chat_completion = lambda **_kwargs: "unused",
+                generate_chat_completion_with_tools = _generate,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setenv(ADMISSION_QUEUE_TIMEOUT_ENV, "0.01")
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+            monkeypatch.setattr(inf_mod, "_select_request_tools", fake_select_tools)
+
+            queue = get_llama_admission_queue("http://llama.tool.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+                enable_tools = True,
+            )
+            try:
+                with pytest.raises(HTTPException) as exc:
+                    await openai_chat_completions(
+                        payload,
+                        request = Request(),
+                        current_subject = "test",
+                    )
+                assert exc.value.status_code == 503
+            finally:
+                blocker.release()
+
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+
+        asyncio.run(_run())
+
+    def test_gguf_tool_non_streaming_cancel_drains_worker_before_releasing_slot(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            async def fake_select_tools(*_args, **_kwargs):
+                return [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ]
+
+            started = threading.Event()
+            released = threading.Event()
+
+            def _tools(**kwargs):
+                cancel_event = kwargs["cancel_event"]
+                started.set()
+                while not cancel_event.is_set():
+                    time.sleep(0.005)
+                released.set()
+                yield from ()
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = True,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                base_url = "http://llama.tool.test",
+                effective_parallel_slots = 1,
+                generate_chat_completion = lambda **_kwargs: "unused",
+                generate_chat_completion_with_tools = _tools,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+            monkeypatch.setattr(inf_mod, "_select_request_tools", fake_select_tools)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+                enable_tools = True,
+            )
+            task = asyncio.create_task(
+                openai_chat_completions(
+                    payload,
+                    request = self._Request(),
+                    current_subject = "test",
+                )
+            )
+            assert await asyncio.to_thread(started.wait, 1.0)
+
+            task.cancel()
+            with pytest.raises(asyncio.CancelledError):
+                await asyncio.wait_for(task, timeout = 1.0)
+
+            assert released.is_set()
+            assert get_llama_admission_queue("http://llama.tool.test").snapshot().active == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
     def test_non_streaming_gguf_n_records_all_monitor_replies(self, monkeypatch):
         import routes.inference as inf_mod
 
@@ -1691,6 +3216,58 @@ class TestGgufVisionToolRouting:
         assert entry["completion_tokens"] == 3
         assert monitor.active_count() == 0
 
+    def test_non_streaming_gguf_cancel_drains_worker(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            started = threading.Event()
+            released = threading.Event()
+
+            def _generate(**kwargs):
+                cancel_event = kwargs["cancel_event"]
+                started.set()
+                while not cancel_event.is_set():
+                    time.sleep(0.005)
+                released.set()
+                yield from ()
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                supports_tools = False,
+                _is_audio = False,
+                model_identifier = "test-gguf",
+                context_length = 4096,
+                generate_chat_completion = _generate,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [{"role": "user", "content": "hi"}],
+            )
+            task = asyncio.create_task(
+                openai_chat_completions(
+                    payload,
+                    request = self._Request(),
+                    current_subject = "test",
+                )
+            )
+            assert await asyncio.to_thread(started.wait, 1.0)
+
+            task.cancel()
+            with pytest.raises(asyncio.CancelledError):
+                await asyncio.wait_for(task, timeout = 1.0)
+
+            assert released.is_set()
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
     def test_standard_gguf_merges_system_and_developer_messages(self, monkeypatch):
         import routes.inference as inf_mod
 
@@ -1801,7 +3378,12 @@ class TestApiMonitorProviderAndCompletionStreams:
         async def is_disconnected(self):
             return False
 
-    async def _run_passthrough_stream(self, monkeypatch, lines):
+    async def _run_passthrough_stream(
+        self,
+        monkeypatch,
+        lines,
+        stream_options = None,
+    ):
         import routes.inference as inf_mod
 
         class Request:
@@ -1829,6 +3411,7 @@ class TestApiMonitorProviderAndCompletionStreams:
             model = "default",
             messages = [ChatMessage(role = "user", content = "hi")],
             stream = True,
+            stream_options = stream_options,
             tools = [
                 {
                     "type": "function",
@@ -1913,6 +3496,134 @@ class TestApiMonitorProviderAndCompletionStreams:
 
         asyncio.run(_run())
 
+    def test_passthrough_stream_forwards_backend_auth_headers(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            captured_headers = {}
+
+            async def fake_send(_client, req, *_args, **_kwargs):
+                captured_headers.update(dict(req.headers))
+                return httpx.Response(200, content = b"")
+
+            class Request:
+                async def is_disconnected(self):
+                    return False
+
+            monitor = ApiMonitor(max_entries = 3)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+                tools = [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ],
+            )
+            response = await _openai_passthrough_stream(
+                Request(),
+                threading.Event(),
+                SimpleNamespace(
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    _auth_headers = {"Authorization": "Bearer secret"},
+                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                ),
+                payload,
+                "chatcmpl-test",
+                "chatcmpl-test",
+                monitor_id = monitor_id,
+            )
+            chunks = [
+                chunk.decode() if isinstance(chunk, bytes) else chunk
+                async for chunk in response.body_iterator
+            ]
+
+            assert "data: [DONE]\n\n" in "".join(chunks)
+            assert captured_headers["authorization"] == "Bearer secret"
+            assert captured_headers["connection"] == "close"
+
+        asyncio.run(_run())
+
+    def test_passthrough_stream_keepalive_while_upstream_headers_are_pending(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            gate = asyncio.Event()
+
+            async def fake_send(*_args, **_kwargs):
+                await gate.wait()
+                return httpx.Response(200, content = b"")
+
+            class Request:
+                async def is_disconnected(self):
+                    return False
+
+            monitor = ApiMonitor(max_entries = 3)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
+            monkeypatch.setattr(
+                inf_mod,
+                "_OPENAI_PASSTHROUGH_PENDING_RESPONSE_KEEPALIVE_S",
+                0.01,
+            )
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+            )
+
+            response = await asyncio.wait_for(
+                _openai_passthrough_stream(
+                    Request(),
+                    threading.Event(),
+                    SimpleNamespace(
+                        base_url = "http://llama.test",
+                        context_length = 4096,
+                        _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                    ),
+                    payload,
+                    "chatcmpl-test",
+                    "chatcmpl-test",
+                    monitor_id = monitor_id,
+                ),
+                timeout = 0.2,
+            )
+
+            first = await asyncio.wait_for(response.body_iterator.__anext__(), timeout = 0.2)
+            assert first == ": keep-alive\n\n"
+
+            gate.set()
+            chunks = [
+                chunk.decode() if isinstance(chunk, bytes) else chunk
+                async for chunk in response.body_iterator
+            ]
+            body = "".join(chunks)
+            assert "data: [DONE]\n\n" in body
+
+        asyncio.run(_run())
+
     def test_passthrough_stream_preheader_non_200_in_window(self, monkeypatch):
         async def _run():
             import routes.inference as inf_mod
@@ -2055,6 +3766,7 @@ class TestApiMonitorProviderAndCompletionStreams:
             body = "".join(chunks)
             assert "data:" in body
             assert '"error"' in body
+            assert "data: [DONE]" in body
             [entry] = monitor.snapshot()
             assert entry["status"] == "error"
             assert "bad" in entry["error"]
@@ -2117,7 +3829,13 @@ class TestApiMonitorProviderAndCompletionStreams:
                 async for chunk in response.body_iterator
             ]
             body = "".join(chunks)
-            payload = json.loads(body.removeprefix("data: ").strip())
+            events = [
+                line.removeprefix("data: ")
+                for line in body.splitlines()
+                if line.startswith("data: ")
+            ]
+            assert events[-1] == "[DONE]"
+            payload = json.loads(events[0])
             assert payload["error"]["code"] == "context_length_exceeded"
             assert payload["error"]["param"] == "messages"
             assert isinstance(payload["error"], dict)
@@ -2207,6 +3925,105 @@ class TestApiMonitorProviderAndCompletionStreams:
 
         asyncio.run(_run())
 
+    def test_passthrough_stream_preheader_immediate_context_retry_adopts_delayed_response(
+        self, monkeypatch
+    ):
+        async def _run():
+            import routes.inference as inf_mod
+
+            gate = asyncio.Event()
+            calls = []
+            err_body = json.dumps(
+                {
+                    "error": {
+                        "message": "request (10000 tokens) exceeds the available context size (2048 tokens)",
+                        "n_prompt_tokens": 10000,
+                        "n_ctx": 2048,
+                    }
+                }
+            ).encode("utf-8")
+            ok_lines = [
+                'data: {"id":"chatcmpl-test","object":"chat.completion.chunk","created":1,'
+                '"model":"gguf","choices":[{"index":0,"delta":{"content":"OK"},'
+                '"finish_reason":null}]}',
+                'data: {"id":"chatcmpl-test","object":"chat.completion.chunk","created":1,'
+                '"model":"gguf","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}',
+                "data: [DONE]",
+            ]
+
+            async def fake_send(_client, req, *_args, **_kwargs):
+                calls.append(json.loads(req.content.decode("utf-8")))
+                if len(calls) == 1:
+                    return httpx.Response(400, content = err_body)
+                await gate.wait()
+                return httpx.Response(200, content = b"")
+
+            async def fake_items(*_args, **_kwargs):
+                for line in ok_lines:
+                    yield line
+
+            class Request:
+                async def is_disconnected(self):
+                    return False
+
+            monitor = ApiMonitor(max_entries = 3)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
+            monkeypatch.setattr(inf_mod, "_aiter_llama_stream_items", fake_items)
+
+            messages = [
+                ChatMessage(role = "system", content = "system"),
+                *[
+                    ChatMessage(role = "user", content = f"turn {idx} " + ("x" * 1000))
+                    for idx in range(8)
+                ],
+            ]
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = messages,
+                stream = True,
+                context_overflow = "truncate_middle",
+            )
+            response = await asyncio.wait_for(
+                _openai_passthrough_stream(
+                    Request(),
+                    threading.Event(),
+                    SimpleNamespace(
+                        base_url = "http://llama.test",
+                        context_length = 2048,
+                        _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                    ),
+                    payload,
+                    "chatcmpl-test",
+                    "chatcmpl-test",
+                    monitor_id = monitor_id,
+                ),
+                timeout = 0.2,
+            )
+            assert isinstance(response, _SameTaskStreamingResponse)
+
+            gate.set()
+            chunks = [
+                chunk.decode() if isinstance(chunk, bytes) else chunk
+                async for chunk in response.body_iterator
+            ]
+            body = "".join(chunks)
+
+            assert "OK" in body
+            assert "context_length_exceeded" not in body
+            assert len(calls) == 2
+            assert len(calls[1]["messages"]) < len(calls[0]["messages"])
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "completed"
+
+        asyncio.run(_run())
+
     def test_passthrough_stream_preheader_delayed_request_error_cleans_up(self, monkeypatch):
         async def _run():
             import routes.inference as inf_mod
@@ -2647,6 +4464,150 @@ class TestApiMonitorProviderAndCompletionStreams:
 
         asyncio.run(_run())
 
+    def test_completions_omitted_max_tokens_falls_back_to_context(self, monkeypatch):
+        # With no env knobs set, an omitted max_tokens must forward the
+        # backend's context length, exactly as on main.
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request:
+                state = SimpleNamespace()
+                url = SimpleNamespace(path = "/v1/completions")
+                method = "POST"
+
+                async def json(self):
+                    return {"prompt": "hi", "stream": False}
+
+            captured = []
+
+            class CapturingClient:
+                async def post(self, _url, *, json, **_kwargs):
+                    captured.append(dict(json))
+                    return httpx.Response(
+                        200,
+                        json = {
+                            "id": "cmpl-test",
+                            "choices": [{"text": "ok"}],
+                            "usage": {
+                                "prompt_tokens": 1,
+                                "completion_tokens": 1,
+                                "total_tokens": 2,
+                            },
+                        },
+                    )
+
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "nonstreaming_client", lambda: CapturingClient())
+            monkeypatch.setattr(
+                inf_mod,
+                "get_llama_cpp_backend",
+                lambda: SimpleNamespace(
+                    is_loaded = True,
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    model_identifier = "gguf",
+                ),
+            )
+
+            await openai_completions(Request(), current_subject = "test")
+
+            assert captured[0]["max_tokens"] == 4096
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_completions_forwards_spec_valid_zero_max_tokens(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request:
+                state = SimpleNamespace()
+                url = SimpleNamespace(path = "/v1/completions")
+                method = "POST"
+
+                async def json(self):
+                    return {"prompt": "hi", "stream": False, "max_tokens": 0}
+
+            captured = []
+
+            class CapturingClient:
+                async def post(self, _url, *, json, **_kwargs):
+                    captured.append(dict(json))
+                    return httpx.Response(
+                        200,
+                        json = {
+                            "id": "cmpl-test",
+                            "choices": [{"text": "", "finish_reason": "length"}],
+                            "usage": {
+                                "prompt_tokens": 1,
+                                "completion_tokens": 0,
+                                "total_tokens": 1,
+                            },
+                        },
+                    )
+
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "nonstreaming_client", lambda: CapturingClient())
+            monkeypatch.setattr(
+                inf_mod,
+                "get_llama_cpp_backend",
+                lambda: SimpleNamespace(
+                    is_loaded = True,
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    model_identifier = "gguf",
+                ),
+            )
+
+            await openai_completions(Request(), current_subject = "test")
+
+            assert captured[0]["max_tokens"] == 0
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_completions_rejects_non_integer_max_tokens_before_forwarding(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request:
+                state = SimpleNamespace()
+                url = SimpleNamespace(path = "/v1/completions")
+                method = "POST"
+
+                async def json(self):
+                    return {"prompt": "hi", "stream": False, "max_tokens": "128"}
+
+            class UnusedClient:
+                async def post(self, *_args, **_kwargs):
+                    raise AssertionError("invalid max_tokens must not reach llama-server")
+
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "nonstreaming_client", lambda: UnusedClient())
+            monkeypatch.setattr(
+                inf_mod,
+                "get_llama_cpp_backend",
+                lambda: SimpleNamespace(
+                    is_loaded = True,
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    model_identifier = "gguf",
+                ),
+            )
+
+            with pytest.raises(HTTPException) as exc:
+                await openai_completions(Request(), current_subject = "test")
+
+            assert exc.value.status_code == 400
+            assert exc.value.detail["error"]["param"] == "max_tokens"
+            assert exc.value.detail["error"]["code"] == "invalid_type"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
     def test_monitor_openai_chunk_records_all_choice_replies(self, monkeypatch):
         import routes.inference as inf_mod
 
@@ -2793,6 +4754,8 @@ class TestApiMonitorProviderAndCompletionStreams:
                 yield 'data: {"choices":[{"delta":{"content":"hello"}}]}'
                 await asyncio.sleep(3600)
 
+            cancel_id = "passthrough-stream-delete-cancel"
+
             monitor = ApiMonitor(max_entries = 3)
             monkeypatch.setattr(inf_mod, "api_monitor", monitor)
             monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
@@ -2807,6 +4770,7 @@ class TestApiMonitorProviderAndCompletionStreams:
                 model = "default",
                 messages = [ChatMessage(role = "user", content = "hi")],
                 stream = True,
+                cancel_id = cancel_id,
                 tools = [
                     {
                         "type": "function",
@@ -2824,6 +4788,7 @@ class TestApiMonitorProviderAndCompletionStreams:
                 SimpleNamespace(
                     base_url = "http://llama.test",
                     context_length = 4096,
+                    _auth_headers = {"Authorization": "Bearer secret"},
                     _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
                 ),
                 payload,
@@ -2835,6 +4800,7 @@ class TestApiMonitorProviderAndCompletionStreams:
             iterator = response.body_iterator
             first = await anext(iterator)
             assert "hello" in first
+            assert cancel_id in inf_mod._CANCEL_REGISTRY
 
             pending = asyncio.create_task(anext(iterator))
             await asyncio.sleep(0)
@@ -2846,6 +4812,271 @@ class TestApiMonitorProviderAndCompletionStreams:
             assert entry["status"] == "cancelled"
             assert entry["reply"] == "hello"
             assert monitor.active_count() == 0
+            assert cancel_id not in inf_mod._CANCEL_REGISTRY
+
+        asyncio.run(_run())
+
+    def test_passthrough_stream_immediate_task_cancel_releases_admission_and_tracker(
+        self, monkeypatch
+    ):
+        async def _run():
+            import routes.inference as inf_mod
+
+            async def fake_cancel_check(*_args, **_kwargs):
+                raise asyncio.CancelledError()
+
+            cancel_id = "passthrough-stream-immediate-task-cancel"
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(
+                inf_mod,
+                "_raise_if_openai_admission_cancelled",
+                fake_cancel_check,
+            )
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+                cancel_id = cancel_id,
+            )
+            backend = SimpleNamespace(
+                base_url = "http://llama.test",
+                context_length = 4096,
+                effective_parallel_slots = 1,
+                _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+            )
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+
+            with pytest.raises(asyncio.CancelledError):
+                await _openai_passthrough_stream(
+                    self._Request(),
+                    threading.Event(),
+                    backend,
+                    payload,
+                    "gguf",
+                    "chatcmpl-test",
+                    monitor_id = monitor_id,
+                )
+
+            assert cancel_id not in inf_mod._CANCEL_REGISTRY
+            assert get_llama_admission_queue("http://llama.test").snapshot().active == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_stream_queued_cancel_before_inner_first_chunk_runs_cleanup(
+        self, monkeypatch
+    ):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request(self._Request):
+                app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+
+            body_holder = {}
+            cleanup_called = threading.Event()
+
+            async def fake_admitted(*_args, admission_lease, tracker, **_kwargs):
+                async def cleanup():
+                    admission_lease.release()
+                    tracker.__exit__(None, None, None)
+                    cleanup_called.set()
+
+                class BlockingBody:
+                    def __init__(self):
+                        self.started = threading.Event()
+                        self.closed = False
+
+                    def __aiter__(self):
+                        return self
+
+                    async def __anext__(self):
+                        self.started.set()
+                        await asyncio.sleep(3600)
+                        raise StopAsyncIteration
+
+                    async def aclose(self):
+                        self.closed = True
+                        await cleanup()
+
+                body = BlockingBody()
+                body_holder["body"] = body
+                return _SameTaskStreamingResponse(
+                    body,
+                    media_type = "text/event-stream",
+                    unstarted_cleanup = cleanup,
+                )
+
+            monkeypatch.setenv(ADMISSION_KEEPALIVE_INTERVAL_ENV, "0.01")
+            monkeypatch.setattr(
+                inf_mod,
+                "_openai_passthrough_stream_admitted",
+                fake_admitted,
+            )
+
+            queue = get_llama_admission_queue("http://llama.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+
+            cancel_id = "queued-inner-unstarted-cleanup"
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+                cancel_id = cancel_id,
+            )
+            response = await _openai_passthrough_stream(
+                Request(),
+                threading.Event(),
+                SimpleNamespace(
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    effective_parallel_slots = 1,
+                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                ),
+                payload,
+                "gguf",
+                "chatcmpl-test",
+            )
+            iterator = response.body_iterator
+            try:
+                chunk = await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+                assert chunk == ": keep-alive\n\n"
+                assert cancel_id in inf_mod._CANCEL_REGISTRY
+
+                blocker.release()
+                pending = asyncio.create_task(iterator.__anext__())
+                for _ in range(100):
+                    if "body" in body_holder:
+                        break
+                    await asyncio.sleep(0.01)
+                body = body_holder["body"]
+                assert await asyncio.to_thread(body.started.wait, 1.0)
+
+                pending.cancel()
+                with pytest.raises(asyncio.CancelledError):
+                    await asyncio.wait_for(pending, timeout = 1.0)
+            finally:
+                aclose = getattr(iterator, "aclose", None)
+                if aclose is not None:
+                    await aclose()
+                blocker.release()
+
+            assert body_holder["body"].closed
+            assert cleanup_called.is_set()
+            assert cancel_id not in inf_mod._CANCEL_REGISTRY
+            assert queue.snapshot().active == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_stream_queued_cancel_after_inner_first_chunk_finalizes_monitor(
+        self, monkeypatch
+    ):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request(self._Request):
+                app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+
+            async def fake_admitted(
+                *_args,
+                monitor_id = None,
+                admission_lease,
+                tracker,
+                **_kwargs,
+            ):
+                async def cleanup():
+                    admission_lease.release()
+                    tracker.__exit__(None, None, None)
+
+                async def body():
+                    try:
+                        yield 'data: {"choices":[{"delta":{"content":"hello"}}]}\n\n'
+                        await asyncio.sleep(3600)
+                    except asyncio.CancelledError:
+                        inf_mod.api_monitor.finish(monitor_id, "cancelled")
+                        raise
+                    finally:
+                        await cleanup()
+
+                return _SameTaskStreamingResponse(
+                    body(),
+                    media_type = "text/event-stream",
+                    unstarted_cleanup = cleanup,
+                )
+
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setenv(ADMISSION_KEEPALIVE_INTERVAL_ENV, "0.01")
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(
+                inf_mod,
+                "_openai_passthrough_stream_admitted",
+                fake_admitted,
+            )
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+
+            queue = get_llama_admission_queue("http://llama.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+
+            cancel_id = "queued-inner-cancel-monitor"
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+                cancel_id = cancel_id,
+            )
+            response = await _openai_passthrough_stream(
+                Request(),
+                threading.Event(),
+                SimpleNamespace(
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    effective_parallel_slots = 1,
+                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                ),
+                payload,
+                "gguf",
+                "chatcmpl-test",
+                monitor_id = monitor_id,
+            )
+            iterator = response.body_iterator
+            try:
+                chunk = await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+                assert chunk == ": keep-alive\n\n"
+
+                blocker.release()
+                first = await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+                assert "hello" in first
+
+                pending = asyncio.create_task(iterator.__anext__())
+                await asyncio.sleep(0)
+                pending.cancel()
+                with pytest.raises(asyncio.CancelledError):
+                    await asyncio.wait_for(pending, timeout = 1.0)
+            finally:
+                aclose = getattr(iterator, "aclose", None)
+                if aclose is not None:
+                    await aclose()
+                blocker.release()
+
+            assert cancel_id not in inf_mod._CANCEL_REGISTRY
+            assert queue.snapshot().active == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
 
         asyncio.run(_run())
 
@@ -2931,6 +5162,313 @@ class TestApiMonitorProviderAndCompletionStreams:
 
         asyncio.run(_run())
 
+    def test_passthrough_usage_done_are_separate_sse_events(self, monkeypatch):
+        async def _run():
+            result = await self._run_passthrough_stream(
+                monkeypatch,
+                [
+                    'data: {"id":"chatcmpl-test","object":"chat.completion.chunk","created":1,"model":"m","choices":[{"index":0,"delta":{"role":"assistant"},"finish_reason":null}]}',
+                    'data: {"id":"chatcmpl-test","object":"chat.completion.chunk","created":1,"model":"m","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}',
+                    'data: {"id":"chatcmpl-test","object":"chat.completion.chunk","created":1,"model":"m","choices":[],"usage":{"prompt_tokens":1,"completion_tokens":1,"total_tokens":2}}',
+                ],
+                stream_options = {"include_usage": True},
+            )
+
+            assert (
+                '"usage":{"prompt_tokens":1,"completion_tokens":1,"total_tokens":2' in result.body
+            )
+            assert "data: [DONE]" in result.body
+            assert "}\n\ndata: [DONE]\n\n" in result.body
+            assert "}\ndata: [DONE]\n\n" not in result.body
+
+        asyncio.run(_run())
+
+    def test_passthrough_stream_queued_request_sends_keepalive_before_upstream(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request:
+                app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+                url = SimpleNamespace(path = "/v1/chat/completions")
+
+                async def is_disconnected(self):
+                    return False
+
+            async def fail_admitted(*_args, **_kwargs):
+                raise AssertionError("upstream must not start while request is queued")
+
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setenv(ADMISSION_KEEPALIVE_INTERVAL_ENV, "0.01")
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_openai_passthrough_stream_admitted", fail_admitted)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+
+            queue = get_llama_admission_queue("http://llama.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+            )
+            response = await _openai_passthrough_stream(
+                Request(),
+                threading.Event(),
+                SimpleNamespace(
+                    base_url = "http://llama.test",
+                    effective_parallel_slots = 1,
+                    context_length = 4096,
+                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                ),
+                payload,
+                "gguf",
+                "chatcmpl-test",
+                monitor_id = monitor_id,
+            )
+            iterator = response.body_iterator
+            try:
+                chunk = await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+                assert chunk == ": keep-alive\n\n"
+                snapshot = queue.snapshot()
+                assert snapshot.active == 1
+                assert snapshot.queued == 1
+            finally:
+                aclose = getattr(iterator, "aclose", None)
+                if aclose is not None:
+                    await aclose()
+                blocker.release()
+
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_non_streaming_admission_timeout_before_upstream(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request:
+                app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+                url = SimpleNamespace(path = "/v1/chat/completions")
+
+                async def is_disconnected(self):
+                    return False
+
+            async def fail_upstream(*_args, **_kwargs):
+                raise AssertionError("upstream must not start while request is queued")
+
+            monkeypatch.setenv(ADMISSION_QUEUE_TIMEOUT_ENV, "0.01")
+            monkeypatch.setattr(
+                inf_mod,
+                "_openai_passthrough_non_streaming_upstream",
+                fail_upstream,
+            )
+
+            queue = get_llama_admission_queue("http://llama.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+            )
+            try:
+                with pytest.raises(HTTPException) as exc:
+                    await _openai_passthrough_non_streaming(
+                        SimpleNamespace(
+                            base_url = "http://llama.test",
+                            effective_parallel_slots = 1,
+                            context_length = 4096,
+                            _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                        ),
+                        payload,
+                        "gguf",
+                        request = Request(),
+                        cancel_event = threading.Event(),
+                    )
+                assert exc.value.status_code == 503
+            finally:
+                blocker.release()
+
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_non_streaming_admission_queue_full_before_upstream(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request:
+                app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+                url = SimpleNamespace(path = "/v1/chat/completions")
+
+                async def is_disconnected(self):
+                    return False
+
+            async def fail_upstream(*_args, **_kwargs):
+                raise AssertionError("upstream must not start when admission queue is full")
+
+            monkeypatch.setenv(ADMISSION_MAX_QUEUE_ENV, "1")
+            monkeypatch.setattr(
+                inf_mod,
+                "_openai_passthrough_non_streaming_upstream",
+                fail_upstream,
+            )
+
+            queue = get_llama_admission_queue("http://llama.test")
+            blocker = queue.reserve(
+                capacity = 1,
+                config = LlamaAdmissionConfig(max_queue = 1),
+            ).lease_nowait()
+            queued = queue.reserve(capacity = 1, config = LlamaAdmissionConfig(max_queue = 1))
+            assert blocker is not None
+            assert queued.lease_nowait() is None
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+            )
+            try:
+                with pytest.raises(HTTPException) as exc:
+                    await _openai_passthrough_non_streaming(
+                        SimpleNamespace(
+                            base_url = "http://llama.test",
+                            effective_parallel_slots = 1,
+                            context_length = 4096,
+                            _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                        ),
+                        payload,
+                        "gguf",
+                        request = Request(),
+                        cancel_event = threading.Event(),
+                    )
+                assert exc.value.status_code == 429
+            finally:
+                queued.cancel()
+                blocker.release()
+
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_non_streaming_immediate_cancel_stops_before_upstream(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            async def fail_upstream(*_args, **_kwargs):
+                raise AssertionError("upstream must not start after client cancellation")
+
+            monkeypatch.setattr(
+                inf_mod,
+                "_openai_passthrough_non_streaming_upstream",
+                fail_upstream,
+            )
+
+            cancel_event = threading.Event()
+            cancel_event.set()
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+            )
+
+            with pytest.raises(HTTPException) as exc:
+                await _openai_passthrough_non_streaming(
+                    SimpleNamespace(
+                        base_url = "http://llama.test",
+                        effective_parallel_slots = 1,
+                        context_length = 4096,
+                        _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                    ),
+                    payload,
+                    "gguf",
+                    monitor_id = monitor_id,
+                    cancel_event = cancel_event,
+                )
+
+            assert exc.value.status_code == 499
+            assert get_llama_admission_queue("http://llama.test").snapshot().active == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_non_streaming_admission_task_cancel_finalizes_monitor(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            async def fake_wait(*_args, **_kwargs):
+                raise asyncio.CancelledError()
+
+            async def fail_upstream(*_args, **_kwargs):
+                raise AssertionError("upstream must not start after admission task cancel")
+
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(
+                inf_mod,
+                "_wait_for_openai_admission_non_streaming",
+                fake_wait,
+            )
+            monkeypatch.setattr(
+                inf_mod,
+                "_openai_passthrough_non_streaming_upstream",
+                fail_upstream,
+            )
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+            )
+
+            with pytest.raises(asyncio.CancelledError):
+                await _openai_passthrough_non_streaming(
+                    SimpleNamespace(
+                        base_url = "http://llama.test",
+                        effective_parallel_slots = 1,
+                        context_length = 4096,
+                        _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                    ),
+                    payload,
+                    "gguf",
+                    monitor_id = monitor_id,
+                    cancel_event = threading.Event(),
+                )
+
+            assert get_llama_admission_queue("http://llama.test").snapshot().active == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
     def test_passthrough_non_streaming_cancel_finalizes_monitor(self, monkeypatch):
         async def _run():
             import routes.inference as inf_mod
@@ -2990,6 +5528,407 @@ class TestApiMonitorProviderAndCompletionStreams:
 
         asyncio.run(_run())
 
+    def test_passthrough_non_streaming_cancel_closes_blocked_upstream_post(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class HangingCancelableClient:
+                def __init__(self):
+                    self.started = asyncio.Event()
+                    self.closed = asyncio.Event()
+
+                async def post(self, *_args, **_kwargs):
+                    self.started.set()
+                    await self.closed.wait()
+                    raise httpx.ReadError("client closed")
+
+                async def aclose(self):
+                    self.closed.set()
+
+            class Request:
+                async def is_disconnected(self):
+                    return False
+
+            client = HangingCancelableClient()
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(
+                inf_mod,
+                "_cancelable_nonstreaming_client",
+                lambda: client,
+            )
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            cancel_event = threading.Event()
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                tools = [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ],
+            )
+
+            task = asyncio.create_task(
+                _openai_passthrough_non_streaming(
+                    SimpleNamespace(
+                        base_url = "http://llama.test",
+                        context_length = 4096,
+                        _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                    ),
+                    payload,
+                    "gguf",
+                    monitor_id = monitor_id,
+                    request = Request(),
+                    cancel_event = cancel_event,
+                )
+            )
+            await asyncio.wait_for(client.started.wait(), 0.2)
+            cancel_event.set()
+
+            with pytest.raises(asyncio.CancelledError):
+                await asyncio.wait_for(task, 0.5)
+
+            assert client.closed.is_set()
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_non_streaming_route_registers_cancel_id(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class HangingCancelableClient:
+                def __init__(self):
+                    self.started = asyncio.Event()
+                    self.closed = asyncio.Event()
+
+                async def post(self, *_args, **_kwargs):
+                    self.started.set()
+                    await self.closed.wait()
+                    raise httpx.ReadError("client closed")
+
+                async def aclose(self):
+                    self.closed.set()
+
+            class Request:
+                state = SimpleNamespace()
+                url = SimpleNamespace(path = "/v1/chat/completions")
+                method = "POST"
+
+                async def is_disconnected(self):
+                    return False
+
+            cancel_id = "passthrough-nonstream-cancel-id"
+            client = HangingCancelableClient()
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_cancelable_nonstreaming_client", lambda: client)
+
+            def _plain(**_kwargs):
+                raise AssertionError("plain GGUF path should not be used")
+
+            monkeypatch.setattr(
+                inf_mod,
+                "get_llama_cpp_backend",
+                lambda: SimpleNamespace(
+                    is_loaded = True,
+                    is_vision = False,
+                    supports_tools = True,
+                    _is_audio = False,
+                    model_identifier = "test-gguf",
+                    context_length = 4096,
+                    base_url = "http://llama.test",
+                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                    generate_chat_completion = _plain,
+                ),
+            )
+
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                cancel_id = cancel_id,
+                tools = [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ],
+            )
+
+            task = asyncio.create_task(
+                openai_chat_completions(
+                    payload,
+                    request = Request(),
+                    current_subject = "test",
+                )
+            )
+            await asyncio.wait_for(client.started.wait(), 0.2)
+            assert cancel_id in inf_mod._CANCEL_REGISTRY
+            assert inf_mod._cancel_by_cancel_id_or_stash(cancel_id) == 1
+
+            with pytest.raises(asyncio.CancelledError):
+                await asyncio.wait_for(task, 0.5)
+
+            assert client.closed.is_set()
+            assert cancel_id not in inf_mod._CANCEL_REGISTRY
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_non_streaming_disconnect_closes_blocked_upstream_post(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class HangingCancelableClient:
+                def __init__(self):
+                    self.started = asyncio.Event()
+                    self.closed = asyncio.Event()
+
+                async def post(self, *_args, **_kwargs):
+                    self.started.set()
+                    await self.closed.wait()
+                    raise httpx.ReadError("client closed")
+
+                async def aclose(self):
+                    self.closed.set()
+
+            class Request:
+                def __init__(self):
+                    self.disconnected = False
+
+                async def is_disconnected(self):
+                    return self.disconnected
+
+            client = HangingCancelableClient()
+            request = Request()
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(
+                inf_mod,
+                "_cancelable_nonstreaming_client",
+                lambda: client,
+            )
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            cancel_event = threading.Event()
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                tools = [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ],
+            )
+
+            task = asyncio.create_task(
+                _openai_passthrough_non_streaming(
+                    SimpleNamespace(
+                        base_url = "http://llama.test",
+                        context_length = 4096,
+                        _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                    ),
+                    payload,
+                    "gguf",
+                    monitor_id = monitor_id,
+                    request = request,
+                    cancel_event = cancel_event,
+                )
+            )
+            await asyncio.wait_for(client.started.wait(), 0.2)
+            request.disconnected = True
+
+            with pytest.raises(asyncio.CancelledError):
+                await asyncio.wait_for(task, 0.5)
+
+            assert client.closed.is_set()
+            assert cancel_event.is_set()
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_non_streaming_forwards_backend_auth_headers(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            captured = {}
+
+            class FakeNonStreamingClient:
+                async def post(self, *_args, **kwargs):
+                    captured["headers"] = kwargs.get("headers")
+                    return httpx.Response(
+                        200,
+                        json = {
+                            "id": "chatcmpl-test",
+                            "object": "chat.completion",
+                            "created": 123,
+                            "model": "gguf",
+                            "choices": [
+                                {
+                                    "index": 0,
+                                    "message": {"role": "assistant", "content": "OK"},
+                                    "finish_reason": "stop",
+                                }
+                            ],
+                        },
+                    )
+
+            monitor = ApiMonitor(max_entries = 3)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(
+                inf_mod,
+                "nonstreaming_client",
+                lambda: FakeNonStreamingClient(),
+            )
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                tools = [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ],
+            )
+
+            response = await _openai_passthrough_non_streaming(
+                SimpleNamespace(
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    _auth_headers = {"Authorization": "Bearer secret"},
+                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                ),
+                payload,
+                "gguf",
+                monitor_id = monitor_id,
+            )
+
+            assert json.loads(response.body)["choices"][0]["message"]["content"] == "OK"
+            assert captured["headers"]["Authorization"] == "Bearer secret"
+            assert captured["headers"]["Connection"] == "close"
+
+        asyncio.run(_run())
+
+    def test_passthrough_non_streaming_forces_upstream_stream_false(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            captured = {}
+
+            class FakeNonStreamingClient:
+                async def __aenter__(self):
+                    return self
+
+                async def __aexit__(self, *_args):
+                    return False
+
+                async def post(self, *_args, **kwargs):
+                    captured["json"] = kwargs.get("json")
+                    return httpx.Response(
+                        200,
+                        json = {
+                            "id": "chatcmpl-test",
+                            "object": "chat.completion",
+                            "created": 123,
+                            "model": "gguf",
+                            "choices": [
+                                {
+                                    "index": 0,
+                                    "message": {"role": "assistant", "content": "OK"},
+                                    "finish_reason": "stop",
+                                }
+                            ],
+                            "usage": {
+                                "prompt_tokens": 1,
+                                "completion_tokens": 1,
+                                "total_tokens": 2,
+                            },
+                        },
+                    )
+
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(
+                inf_mod,
+                "nonstreaming_client",
+                lambda: FakeNonStreamingClient(),
+            )
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+                stream_options = {"include_usage": True},
+                tools = [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ],
+            )
+
+            await _openai_passthrough_non_streaming(
+                SimpleNamespace(
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                ),
+                payload,
+                "gguf",
+                monitor_id = monitor_id,
+            )
+
+            assert captured["json"]["stream"] is False
+            assert "stream_options" not in captured["json"]
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "completed"
+
+        asyncio.run(_run())
+
     def test_passthrough_clean_eof_finalizes_monitor(self, monkeypatch):
         async def _run():
             result = await self._run_passthrough_stream(
@@ -3009,6 +5948,216 @@ class TestApiMonitorProviderAndCompletionStreams:
 
         asyncio.run(_run())
 
+    def test_passthrough_finish_without_done_closes_stream_early(self, monkeypatch):
+        # Some llama-server builds emit the finish chunk and then hold the HTTP
+        # stream open without sending [DONE]; the terminal classifier must end
+        # the client stream promptly instead of hanging on the open socket.
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request:
+                async def is_disconnected(self):
+                    return False
+
+            async def fake_send(*_args, **_kwargs):
+                return httpx.Response(200, content = b"")
+
+            async def fake_items(*_args, **_kwargs):
+                yield 'data: {"choices":[{"index":0,"delta":{"content":"hi"},"finish_reason":null}]}'
+                yield 'data: {"choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}'
+                await asyncio.Event().wait()  # upstream never closes
+
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
+            monkeypatch.setattr(inf_mod, "_aiter_llama_stream_items", fake_items)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+                tools = [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ],
+            )
+
+            response = await _openai_passthrough_stream(
+                Request(),
+                threading.Event(),
+                SimpleNamespace(
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                ),
+                payload,
+                "gguf",
+                "chatcmpl-test",
+                monitor_id = monitor_id,
+            )
+
+            async def _consume():
+                return [chunk async for chunk in response.body_iterator]
+
+            chunks = await asyncio.wait_for(_consume(), timeout = 2)
+            body = "".join(chunks)
+
+            assert '"finish_reason":"stop"' in body.replace(" ", "")
+            assert body.endswith("data: [DONE]\n\n")
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "completed"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_stall_after_finish_closes_cleanly(self, monkeypatch):
+        # include_usage keeps the stream open past the finish chunk waiting for
+        # the usage chunk; if that never arrives, the post-terminal grace path
+        # must close with a clean [DONE], not an in-band error.
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request:
+                async def is_disconnected(self):
+                    return False
+
+            async def fake_send(*_args, **_kwargs):
+                return httpx.Response(200, content = b"")
+
+            async def fake_items(*_args, **_kwargs):
+                yield 'data: {"choices":[{"index":0,"delta":{"content":"hi"},"finish_reason":null}]}'
+                yield 'data: {"choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}'
+                raise httpx.ReadTimeout("usage chunk never arrived")
+
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
+            monkeypatch.setattr(inf_mod, "_aiter_llama_stream_items", fake_items)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+                stream_options = {"include_usage": True},
+                tools = [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ],
+            )
+
+            response = await _openai_passthrough_stream(
+                Request(),
+                threading.Event(),
+                SimpleNamespace(
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                ),
+                payload,
+                "gguf",
+                "chatcmpl-test",
+                monitor_id = monitor_id,
+            )
+            chunks = [chunk async for chunk in response.body_iterator]
+            body = "".join(chunks)
+
+            assert '"type":"api_error"' not in body.replace(" ", "")
+            assert body.endswith("data: [DONE]\n\n")
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "completed"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_stream_stall_after_data_emits_error(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            class Request:
+                async def is_disconnected(self):
+                    return False
+
+            async def fake_send(*_args, **_kwargs):
+                return httpx.Response(200, content = b"")
+
+            async def fake_items(*_args, **_kwargs):
+                yield 'data: {"choices":[{"delta":{"content":"hello"}}]}'
+                raise httpx.ReadTimeout("upstream went silent")
+
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
+            monkeypatch.setattr(inf_mod, "_aiter_llama_stream_items", fake_items)
+            monitor_id = monitor.start(
+                endpoint = "/v1/chat/completions",
+                method = "POST",
+                model = "gguf",
+                prompt = "hi",
+            )
+            payload = ChatCompletionRequest(
+                model = "default",
+                messages = [ChatMessage(role = "user", content = "hi")],
+                stream = True,
+                tools = [
+                    {
+                        "type": "function",
+                        "function": {
+                            "name": "lookup",
+                            "parameters": {"type": "object", "properties": {}},
+                        },
+                    }
+                ],
+            )
+
+            response = await _openai_passthrough_stream(
+                Request(),
+                threading.Event(),
+                SimpleNamespace(
+                    base_url = "http://llama.test",
+                    context_length = 4096,
+                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+                ),
+                payload,
+                "gguf",
+                "chatcmpl-test",
+                monitor_id = monitor_id,
+            )
+            chunks = [chunk async for chunk in response.body_iterator]
+            body = "".join(chunks)
+
+            assert 'data: {"choices":[{"delta":{"content":"hello"}}]}\n\n' in body
+            assert '"finish_reason"' not in body.replace(" ", "")
+            assert '"type":"api_error"' in body.replace(" ", "")
+            assert "still processing the prompt" in body
+            assert body.endswith("data: [DONE]\n\n")
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "error"
+            assert "still processing the prompt" in entry["error"]
+            assert entry["reply"] == "hello"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
 
 class TestApiMonitorSafetensorsUsage:
     class _Request:
@@ -3512,6 +6661,15 @@ class TestApiMonitorAudioInput:
 class TestResponsesChatTemplateKwargs:
     _messages = [ChatMessage(role = "user", content = "What is 100 - 67?")]
 
+    class _Request:
+        app = SimpleNamespace(state = SimpleNamespace(llama_parallel_slots = 1))
+        state = SimpleNamespace()
+        url = SimpleNamespace(path = "/v1/responses")
+        method = "POST"
+
+        async def is_disconnected(self):
+            return False
+
     def test_enable_thinking_lifted_from_extra_body(self):
         payload = ResponsesRequest(
             model = "qwen-local",
@@ -3544,6 +6702,113 @@ class TestResponsesChatTemplateKwargs:
         chat_req = _build_chat_request(payload, self._messages, stream = False)
         assert chat_req.enable_thinking is None
 
+    def test_responses_stream_queued_request_sends_keepalive_before_upstream(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            async def fail_send(*_args, **_kwargs):
+                raise AssertionError("responses upstream must not start while queued")
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                base_url = "http://llama.responses.test",
+                context_length = 4096,
+                effective_parallel_slots = 1,
+                _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setenv(ADMISSION_KEEPALIVE_INTERVAL_ENV, "0.01")
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fail_send)
+
+            queue = get_llama_admission_queue("http://llama.responses.test")
+            blocker = queue.reserve(capacity = 1, config = LlamaAdmissionConfig()).lease_nowait()
+            assert blocker is not None
+            monitor_id = monitor.start(
+                endpoint = "/v1/responses",
+                method = "POST",
+                model = "qwen-local",
+                prompt = "hi",
+            )
+            payload = ResponsesRequest(model = "qwen-local", input = "hi", stream = True)
+
+            response = await _responses_stream(
+                payload,
+                [ChatMessage(role = "user", content = "hi")],
+                self._Request(),
+                monitor_id,
+            )
+            iterator = response.body_iterator
+            try:
+                chunk = await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+                assert chunk == ": keep-alive\n\n"
+                snapshot = queue.snapshot()
+                assert snapshot.active == 1
+                assert snapshot.queued == 1
+            finally:
+                aclose = getattr(iterator, "aclose", None)
+                if aclose is not None:
+                    await aclose()
+                blocker.release()
+
+            snapshot = queue.snapshot()
+            assert snapshot.active == 0
+            assert snapshot.queued == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_responses_stream_cancel_after_created_finalizes_monitor_and_slot(self, monkeypatch):
+        async def _run():
+            import routes.inference as inf_mod
+
+            async def fail_send(*_args, **_kwargs):
+                raise AssertionError("responses upstream must not start after created cancel")
+
+            backend = SimpleNamespace(
+                is_loaded = True,
+                is_vision = False,
+                base_url = "http://llama.responses.test",
+                context_length = 4096,
+                effective_parallel_slots = 1,
+                _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+            )
+            monitor = ApiMonitor(max_entries = 3)
+            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+            monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fail_send)
+            monitor_id = monitor.start(
+                endpoint = "/v1/responses",
+                method = "POST",
+                model = "qwen-local",
+                prompt = "hi",
+            )
+            payload = ResponsesRequest(model = "qwen-local", input = "hi", stream = True)
+
+            response = await _responses_stream(
+                payload,
+                [ChatMessage(role = "user", content = "hi")],
+                self._Request(),
+                monitor_id,
+            )
+            iterator = response.body_iterator
+            first = await asyncio.wait_for(iterator.__anext__(), timeout = 0.2)
+            assert "event: response.created" in first
+
+            with pytest.raises(asyncio.CancelledError):
+                await iterator.athrow(asyncio.CancelledError())
+
+            assert get_llama_admission_queue("http://llama.responses.test").snapshot().active == 0
+            [entry] = monitor.snapshot()
+            assert entry["status"] == "cancelled"
+            assert monitor.active_count() == 0
+
+        asyncio.run(_run())
+
 
 # =====================================================================
 # GGUF chat-template role alternation: coalesce orphaned user turns left
diff --git a/studio/backend/tests/test_passthrough_healing.py b/studio/backend/tests/test_passthrough_healing.py
index 83bcc5864a..9b22ab8b05 100644
--- a/studio/backend/tests/test_passthrough_healing.py
+++ b/studio/backend/tests/test_passthrough_healing.py
@@ -515,6 +515,7 @@ class ScriptedClient:
         _url,
         json = None,
         timeout = None,
+        headers = None,
     ):
         self.posts.append(json)
         return httpx.Response(200, json = self.bodies[min(len(self.posts) - 1, len(self.bodies) - 1)])
diff --git a/studio/backend/tests/test_response_template_markers.py b/studio/backend/tests/test_response_template_markers.py
new file mode 100644
index 0000000000..8c813e62f2
--- /dev/null
+++ b/studio/backend/tests/test_response_template_markers.py
@@ -0,0 +1,216 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""TEMPLATE_TO_RESPONSES_MAPPER markers must match what the templates render.
+
+The manual instruction/response markers are the fallback for
+train_on_completions when auto-detection is unavailable, so a marker that
+never matches the rendered chat template masks every assistant token and the
+run dies on the all-labels-masked safety net. Six template families shipped
+such markers:
+
+  mistral           - "[INST] " / " [/INST]": the surrounding spaces fold into
+                      the neighbouring tokens ("[INST]" is a single special
+                      token in Mistral v0.3), so the padded strings never match.
+  llama             - same space folding, plus llama-2 tokenizes [INST] after
+                      <s> as bare "[" on transformers 5.x while the standalone
+                      encoding gives "▁[", so the marker must anchor on <s>.
+  starling          - trailing space after "GPT4 Correct Assistant:" folds
+                      into the next content token ("▁Hello").
+  glm               - "[gMASK]<sop>" renders once at text start, never before
+                      later user turns; "<think>" is generation scaffolding
+                      that non-final turns render as a lone "</think>".
+  qwen3-thinking    - "<think>" is stripped from non-final assistant turns
+                      (Qwen3-Thinking-2507) or never rendered (QwQ).
+  zephyr            - role tags are plain text, and SentencePiece tokenizes
+                      "<|assistant|>" differently at text start than after
+                      "</s>\\n" mid-conversation; the markers need the leading
+                      newline anchor to tokenize like a real turn boundary.
+
+Literal assertions run everywhere; the token-level masking checks need the
+representative tokenizers plus unsloth_zoo and skip when either is
+unavailable (offline CI).
+"""
+
+from __future__ import annotations
+
+import importlib.util
+import sys
+from pathlib import Path
+
+import pytest
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+# model_mappings is dependency-free: load it directly so these tests run
+# without the studio venv / package import side effects.
+_MM_PATH = Path(_BACKEND_DIR) / "utils" / "datasets" / "model_mappings.py"
+_mm_spec = importlib.util.spec_from_file_location("_marker_test_mm", _MM_PATH)
+model_mappings = importlib.util.module_from_spec(_mm_spec)
+_mm_spec.loader.exec_module(model_mappings)
+
+T2R = model_mappings.TEMPLATE_TO_RESPONSES_MAPPER
+
+
+# ── Fixed entries: markers derived from what each representative tokenizer
+#    actually renders (see PR for the token-level derivation). ──
+EXPECTED_FIXED = {
+    "mistral": {"instruction": "[INST]", "response": "[/INST]"},
+    "llama": {"instruction": "<s>[INST]", "response": "[/INST]"},
+    "starling": {"instruction": "GPT4 Correct User:", "response": "GPT4 Correct Assistant:"},
+    "glm": {"instruction": "<|user|>", "response": "<|assistant|>"},
+    "qwen3-thinking": {"instruction": "<|im_start|>user\n", "response": "<|im_start|>assistant\n"},
+    "zephyr": {"instruction": "\n<|user|>\n", "response": "\n<|assistant|>\n"},
+}
+
+# Spot-pin some known-good entries so a refactor cannot silently change them.
+EXPECTED_UNCHANGED = {
+    "qwen3": {"instruction": "<|im_start|>user\n", "response": "<|im_start|>assistant\n"},
+    "llama-3.1": {
+        "instruction": "<|start_header_id|>user<|end_header_id|>\n\n",
+        "response": "<|start_header_id|>assistant<|end_header_id|>\n\n",
+    },
+    "phi-4": {
+        "instruction": "<|im_start|>user<|im_sep|>",
+        "response": "<|im_start|>assistant<|im_sep|>",
+    },
+    "gemma-3": {"instruction": "<start_of_turn>user\n", "response": "<start_of_turn>model\n"},
+    "gpt-oss": {
+        "instruction": "<|start|>user<|message|>",
+        "response": "<|start|>assistant<|channel|>final<|message|>",
+    },
+}
+
+
+@pytest.mark.parametrize("template", sorted(EXPECTED_FIXED))
+def test_fixed_marker_literals(template):
+    assert T2R[template] == EXPECTED_FIXED[template]
+
+
+@pytest.mark.parametrize("template", sorted(EXPECTED_UNCHANGED))
+def test_unchanged_marker_literals(template):
+    assert T2R[template] == EXPECTED_UNCHANGED[template]
+
+
+def test_no_marker_is_empty_or_whitespace():
+    for template, parts in T2R.items():
+        assert parts["instruction"].strip(), template
+        assert parts["response"].strip(), template
+
+
+# ── Token-level checks: markers must select exactly the assistant turns on a
+#    rendered two-turn fixture, and the final EOS label must never be -100. ──
+
+REPRESENTATIVES = {
+    "mistral": ["unsloth/mistral-7b-instruct-v0.3"],
+    "llama": ["unsloth/llama-2-7b-chat"],
+    "starling": ["unsloth/Starling-LM-7B-beta"],
+    "glm": ["unsloth/GLM-4.7-Flash"],
+    "qwen3-thinking": ["unsloth/Qwen3-4B-Thinking-2507", "Qwen/QwQ-32B"],
+    "zephyr": ["unsloth/zephyr-sft"],
+}
+
+FIXTURE = [
+    {"role": "user", "content": "zebra alpha question one?"},
+    {"role": "assistant", "content": "grape reply number one."},
+    {"role": "user", "content": "zebra beta question two?"},
+    {"role": "assistant", "content": "grape reply number two."},
+]
+
+
+def _load_tokenizer(repo):
+    try:
+        from transformers import AutoTokenizer
+    except Exception as e:  # pragma: no cover
+        pytest.skip(f"transformers unavailable: {e}")
+    try:
+        return AutoTokenizer.from_pretrained(repo)
+    except OSError as e:
+        pytest.skip(f"tokenizer {repo} unavailable (offline?): {e}")
+    except Exception:
+        # Tokenizer class newer than this transformers (e.g. GLM-4.7's
+        # TokenizersBackend): build directly from tokenizer.json.
+        try:
+            import json as _json
+            from huggingface_hub import hf_hub_download
+            from transformers import PreTrainedTokenizerFast
+
+            with open(hf_hub_download(repo, "tokenizer_config.json"), encoding = "utf-8") as f:
+                cfg = _json.load(f)
+            tok_file = hf_hub_download(repo, "tokenizer.json")
+
+            def _tokval(v):
+                return v["content"] if isinstance(v, dict) else v
+
+            return PreTrainedTokenizerFast(
+                tokenizer_file = tok_file,
+                chat_template = cfg.get("chat_template"),
+                **{
+                    k: _tokval(cfg[k])
+                    for k in ("bos_token", "eos_token", "pad_token", "unk_token")
+                    if cfg.get(k) is not None
+                },
+            )
+        except Exception as e:
+            pytest.skip(f"tokenizer {repo} unavailable (offline?): {e}")
+
+
+def _train_on_responses_only():
+    try:
+        from unsloth_zoo.dataset_utils import train_on_responses_only
+    except Exception as e:
+        pytest.skip(f"unsloth_zoo unavailable: {e}")
+    return train_on_responses_only
+
+
+@pytest.mark.parametrize(
+    "template,repo",
+    [(t, r) for t, repos in sorted(REPRESENTATIVES.items()) for r in repos],
+)
+def test_fixed_markers_token_level(template, repo):
+    tor = _train_on_responses_only()
+    tok = _load_tokenizer(repo)
+    parts = T2R[template]
+
+    msgs = [{"role": "system", "content": "You are a terse assistant."}] + FIXTURE
+    try:
+        ids = tok.apply_chat_template(msgs, tokenize = True, add_generation_prompt = False)
+        if hasattr(ids, "keys"):
+            ids = ids["input_ids"]  # transformers 5.x returns a BatchEncoding
+    except Exception:
+        ids = tok.apply_chat_template(FIXTURE, tokenize = True, add_generation_prompt = False)
+        if hasattr(ids, "keys"):
+            ids = ids["input_ids"]
+
+    fn = tor(
+        None,
+        instruction_part = parts["instruction"],
+        response_part = parts["response"],
+        tokenizer = tok,
+        return_function = True,
+    )
+    labels = fn({"input_ids": [list(ids)]})["labels"][0]
+
+    n = len(ids)
+    trained = tok.decode([ids[i] for i in range(n) if labels[i] != -100])
+    masked = tok.decode([ids[i] for i in range(n) if labels[i] == -100])
+
+    # User and system content fully masked
+    assert "question one" not in trained and "question one" in masked
+    assert "question two" not in trained and "question two" in masked
+    assert "terse assistant" not in trained
+    # EVERY assistant turn trained, not just the last
+    assert "reply number one" in trained
+    assert "reply number two" in trained
+    # The final EOS (last non-whitespace token) must never be -100, or the
+    # fine-tuned model never learns to stop generating.
+    i = n - 1
+    while i > 0 and tok.decode([ids[i]]).strip() == "":
+        i -= 1
+    assert labels[i] != -100, f"final token {tok.convert_ids_to_tokens(int(ids[i]))!r} is masked"
+
+
+if __name__ == "__main__":
+    raise SystemExit(pytest.main([__file__, "-v"]))
diff --git a/studio/backend/tests/test_think_prefill_reemit.py b/studio/backend/tests/test_think_prefill_reemit.py
new file mode 100644
index 0000000000..300ff92776
--- /dev/null
+++ b/studio/backend/tests/test_think_prefill_reemit.py
@@ -0,0 +1,89 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""
+Unit tests for detect_think_prefill.
+
+Reasoning templates (Qwen3.6-style) end the generation prompt with an open
+``<think>\\n`` so the model starts reasoning immediately. skip_prompt
+streaming drops that opening tag, so the safetensors/MLX paths must re-emit
+it for the frontend's <think> parser to render a thinking block.
+"""
+
+import os
+import sys
+
+_backend = os.path.join(os.path.dirname(__file__), "..")
+sys.path.insert(0, _backend)
+
+from core.inference.chat_template_helpers import detect_think_prefill
+
+
+QWEN_PROMPT = "<|im_start|>user\nHi!<|im_end|>\n<|im_start|>assistant\n"
+
+
+def test_open_think_prefill_reemitted():
+    """Qwen3.6-style enable_thinking=True prompt tail: <think>\\n."""
+    assert detect_think_prefill(QWEN_PROMPT + "<think>\n") == "<think>\n"
+
+
+def test_bare_open_think_prefill_reemitted():
+    """Prefill without trailing newline still detected."""
+    assert detect_think_prefill(QWEN_PROMPT + "<think>") == "<think>"
+
+
+def test_closed_think_prefill_not_reemitted():
+    """enable_thinking=False prefills a closed, empty think block."""
+    assert detect_think_prefill(QWEN_PROMPT + "<think>\n\n</think>\n\n") == ""
+
+
+def test_prompt_without_think_untouched():
+    """Non-reasoning templates produce no prefix."""
+    assert detect_think_prefill(QWEN_PROMPT) == ""
+
+
+def test_historical_think_blocks_ignored():
+    """A closed think block in a prior assistant turn (preserve_thinking)
+    must not trigger re-emission when the generation tail is plain."""
+    prompt = (
+        "<|im_start|>user\nHi!<|im_end|>\n"
+        "<|im_start|>assistant\n<think>\nprior reasoning\n</think>\n\nHello!<|im_end|>\n"
+        "<|im_start|>user\nAgain?<|im_end|>\n<|im_start|>assistant\n"
+    )
+    assert detect_think_prefill(prompt) == ""
+
+
+def test_historical_blocks_plus_open_prefill():
+    """Prior closed blocks plus a fresh open prefill: only the tail matters."""
+    prompt = (
+        "<|im_start|>assistant\n<think>\nprior\n</think>\n\nHello!<|im_end|>\n"
+        "<|im_start|>assistant\n<think>\n"
+    )
+    assert detect_think_prefill(prompt) == "<think>\n"
+
+
+def test_content_after_open_tag_not_reemitted():
+    """If non-whitespace follows the tag it is not a plain prefill."""
+    assert detect_think_prefill(QWEN_PROMPT + "<think>\npartial reasoning") == ""
+
+
+def test_empty_and_none_prompts():
+    assert detect_think_prefill("") == ""
+    assert detect_think_prefill(None) == ""
+
+
+def test_guard_suppresses_when_close_tag_is_special():
+    """If </think> is a special token, skip_special_tokens strips the model's
+    close tag, so re-emitting the open would leave an unclosed block. Guard off."""
+    specials = ["<|im_end|>", "<think>", "</think>"]
+    assert detect_think_prefill(QWEN_PROMPT + "<think>\n", specials) == ""
+
+
+def test_guard_emits_when_think_not_special():
+    specials = ["<|im_end|>", "<|endoftext|>"]
+    assert detect_think_prefill(QWEN_PROMPT + "<think>\n", specials) == "<think>\n"
+
+
+def test_guard_default_and_empty_keep_emitting():
+    assert detect_think_prefill(QWEN_PROMPT + "<think>\n", None) == "<think>\n"
+    assert detect_think_prefill(QWEN_PROMPT + "<think>\n", []) == "<think>\n"
diff --git a/studio/backend/tests/test_torchao_select.py b/studio/backend/tests/test_torchao_select.py
index 2d3dc5fbff..e4775a10a6 100644
--- a/studio/backend/tests/test_torchao_select.py
+++ b/studio/backend/tests/test_torchao_select.py
@@ -32,16 +32,23 @@ def _load_module(monkeypatch):
 @pytest.mark.parametrize(
     "torch_version, expected",
     [
-        # torch 2.10 (the reported bug: cu130 resolves 2.10.0) -> 0.16.0,
-        # independent of the local +cuXXX/+rocm/+cpu suffix or patch level.
-        ("2.10.0+cu130", "torchao==0.16.0"),
+        # torch 2.10 on CUDA <= 12 -> 0.16.0 (its cpp is built for torch 2.10.0 and
+        # loads against the CUDA-12 PyPI wheel). Independent of patch level.
+        ("2.10.0+cu128", "torchao==0.16.0"),
+        ("2.10.0+cu126", "torchao==0.16.0"),
         ("2.10.0+rocm6.4", "torchao==0.16.0"),
         ("2.10.0+cpu", "torchao==0.16.0"),
         ("2.10.1", "torchao==0.16.0"),
         ("2.10.0", "torchao==0.16.0"),
-        # Pre-release / dev / rc builds: the minor is cleaned of non-digits.
+        # torch 2.10 on CUDA >= 13 (Blackwell / cu130): 0.16.0's CUDA-12 cpp can't
+        # load against a CUDA-13 torch (libcudart.so.12 error), so use 0.17.0.
+        ("2.10.0+cu130", "torchao==0.17.0"),
+        ("2.10.0+cu140", "torchao==0.17.0"),
+        # Pre-release / dev / rc builds: the minor is cleaned of non-digits; the
+        # CUDA tag still decides 0.16.0 vs 0.17.0.
         ("2.10.0rc1", "torchao==0.16.0"),
-        ("2.10.0.dev20250804+cu130", "torchao==0.16.0"),
+        ("2.10.0.dev20250804+cu130", "torchao==0.17.0"),
+        ("2.10.0.dev20250804+cu128", "torchao==0.16.0"),
         ("2.10rc1", "torchao==0.16.0"),
         # torch 2.11 (reachable via ROCm rocm7.2) and forward -> 0.17.0.
         ("2.11.0+cu130", "torchao==0.17.0"),
diff --git a/studio/backend/tests/test_training_preflight.py b/studio/backend/tests/test_training_preflight.py
index 54048a65dd..47c6669f8f 100644
--- a/studio/backend/tests/test_training_preflight.py
+++ b/studio/backend/tests/test_training_preflight.py
@@ -6,9 +6,15 @@ empty-chat-template crash) before train(). The real methods are bound onto a lig
 fake self so the production logic runs against controlled batches."""
 
 import importlib
+import json
+import os
+import queue
+import subprocess
 import sys
+import threading
 import types
 import unittest
+from pathlib import Path
 from types import SimpleNamespace
 from unittest.mock import MagicMock
 
@@ -184,5 +190,231 @@ class TestChatTemplateRendersEmpty(unittest.TestCase):
         self.assertFalse(s._chat_template_renders_empty())
 
 
+def _clear_trainer_module(package: str):
+    sys.modules.pop(f"{package}.trainer", None)
+    pkg = sys.modules.get(package)
+    if pkg is not None and hasattr(pkg, "trainer"):
+        delattr(pkg, "trainer")
+
+
+def _set_training_platform(monkeypatch, package: str, backend: str):
+    training_mod = importlib.import_module(f"{package}.training")
+    from utils.hardware import hardware as hw
+
+    monkeypatch.setattr(hw, "DEVICE", None)
+    monkeypatch.setattr(
+        training_mod.platform,
+        "system",
+        lambda: "Darwin" if backend == "mlx" else "Linux",
+    )
+    monkeypatch.setattr(
+        training_mod.platform,
+        "machine",
+        lambda: "arm64" if backend == "mlx" else "x86_64",
+    )
+
+
+def _load_trainer_module(
+    monkeypatch,
+    backend: str,
+    package: str = "core.training",
+):
+    _set_training_platform(monkeypatch, package, backend)
+    _clear_trainer_module(package)
+    if package in sys.modules:
+        importlib.reload(sys.modules[package])
+    trainer_mod = importlib.import_module(f"{package}.trainer")
+    training_mod = importlib.import_module(f"{package}.training")
+    monkeypatch.setattr(
+        training_mod._MLXTrainerAdapter,
+        "_activate_transformers_for_model",
+        lambda self, model_name, hf_token: None,
+    )
+    return trainer_mod
+
+
+class _ExitedProc:
+    def join(self, timeout = None):
+        return None
+
+    def is_alive(self):
+        return False
+
+
+class _TerminableProc:
+    def __init__(self):
+        self.terminated = False
+        self._done = threading.Event()
+
+    def join(self, timeout = None):
+        self._done.wait(timeout = timeout or 5)
+
+    def is_alive(self):
+        return not self.terminated
+
+    def terminate(self):
+        self.terminated = True
+        self._done.set()
+
+
+def test_unsloth_trainer_dispatches_for_mlx_and_torch(monkeypatch):
+    trainer_mod = _load_trainer_module(monkeypatch, "mlx")
+
+    mlx_trainer = trainer_mod.UnslothTrainer()
+
+    assert type(mlx_trainer).__module__ == "core.training.training"
+    assert mlx_trainer.get_training_progress().status_message == "Ready to train"
+
+    trainer_mod = _load_trainer_module(monkeypatch, "torch")
+
+    assert trainer_mod.UnslothTrainer().__class__ is trainer_mod.UnslothTrainer
+
+
+def test_cli_mlx_trainer_activates_before_importing_trainer():
+    repo_root = Path(__file__).resolve().parents[3]
+    script = """
+import json
+import sys
+import unsloth_cli.commands.train as train_cmd
+from studio.backend.core.training import training as training_mod
+from utils.hardware import hardware as hw
+
+training_mod.platform.system = lambda: "Darwin"
+training_mod.platform.machine = lambda: "arm64"
+hw.DEVICE = None
+events = []
+
+def fake_activate(model_name, hf_token):
+    events.append({
+        "model_name": model_name,
+        "trainer_loaded": "studio.backend.core.training.trainer" in sys.modules,
+    })
+
+train_cmd._activate_mlx_transformers = fake_activate
+trainer = train_cmd._create_cli_trainer("mlx-community/Qwen3-0.6B-4bit", None)
+print(json.dumps({
+    "trainer_module": type(trainer).__module__,
+    "events": events,
+}))
+"""
+    env = os.environ.copy()
+    env["PYTHONPATH"] = os.pathsep.join(
+        [str(repo_root), str(repo_root / "studio" / "backend"), env.get("PYTHONPATH", "")]
+    )
+    result = subprocess.run(
+        [sys.executable, "-c", script],
+        cwd = repo_root,
+        env = env,
+        text = True,
+        stdout = subprocess.PIPE,
+        stderr = subprocess.PIPE,
+        check = True,
+    )
+    payload = json.loads(result.stdout)
+
+    assert payload["trainer_module"] == "studio.backend.core.training.training"
+    assert payload["events"] == [
+        {"model_name": "mlx-community/Qwen3-0.6B-4bit", "trainer_loaded": False}
+    ]
+
+
+def test_mlx_adapter_builds_config_and_reports_completion(tmp_path, monkeypatch):
+    trainer_mod = _load_trainer_module(monkeypatch, "mlx")
+    captured = {}
+
+    def fake_run_worker(config, event_queue, stop_queue):
+        captured["config"] = config
+        event_queue.put({"type": "progress", "step": 1, "total_steps": 1, "loss": 0.25})
+        event_queue.put(
+            {"type": "complete", "status_message": "done", "output_dir": config["output_dir"]}
+        )
+
+    trainer = trainer_mod.UnslothTrainer()
+    monkeypatch.setattr(trainer, "_run_mlx_worker", fake_run_worker)
+
+    assert trainer.load_model("mlx-community/Qwen3-0.6B-4bit", max_seq_length = 1024)
+    assert trainer.prepare_model_for_training(use_lora = False)
+    dataset, eval_dataset = trainer.load_and_format_dataset("org/dataset")
+    output_dir = tmp_path / "mlx-out"
+
+    assert trainer.start_training(
+        dataset = dataset,
+        eval_dataset = eval_dataset,
+        output_dir = output_dir,
+        project_name = "Sales Assistant",
+        max_steps = 1,
+        learning_rate = 3e-4,
+    )
+    trainer.training_thread.join(timeout = 5)
+
+    progress = trainer.get_training_progress()
+    config = captured["config"]
+    assert progress.is_completed
+    assert progress.output_dir == str(output_dir.resolve())
+    progress.status_message = "mutated"
+    assert trainer.get_training_progress().status_message == "done"
+    assert config["model_name"] == "mlx-community/Qwen3-0.6B-4bit"
+    assert config["project_name"] == "Sales Assistant"
+    assert config["hf_dataset"] == "org/dataset"
+    assert config["training_type"] == "Full Finetuning"
+    assert config["load_in_4bit"] is False
+    assert config["max_seq_length"] == 1024
+    assert config["learning_rate"] == 3e-4
+    assert config["output_dir"] == str(output_dir.resolve())
+    assert config["allow_external_output_dir"] is True
+
+
+def test_mlx_worker_helpers_cover_cli_paths(tmp_path, monkeypatch):
+    _load_trainer_module(monkeypatch, "mlx")
+    from core.training.worker import (
+        _resolve_mlx_local_dataset_files,
+        _resolve_mlx_output_dir,
+    )
+
+    dataset = tmp_path / "train.jsonl"
+    dataset.write_text('{"text":"hello"}\n', encoding = "utf-8")
+    monkeypatch.chdir(tmp_path)
+
+    assert _resolve_mlx_local_dataset_files(["train.jsonl"]) == [str(dataset)]
+    assert _resolve_mlx_output_dir(
+        {"output_dir": "cli-out", "allow_external_output_dir": True},
+        "mlx-community/Qwen3-0.6B-4bit",
+    ) == str((tmp_path / "cli-out").resolve())
+
+
+def test_run_mlx_training_process_applies_side_effects_before_hardware_detection(monkeypatch):
+    _load_trainer_module(monkeypatch, "mlx")
+    from core.training import worker
+    from utils.hardware import hardware as hw
+
+    order = []
+
+    def fake_activate(model_name, hf_token):
+        order.append(("activate", model_name, hf_token))
+
+    def fake_detect_hardware():
+        order.append("detect")
+        hw.DEVICE = hw.DeviceType.CPU
+        return hw.DEVICE
+
+    monkeypatch.delenv("HF_HUB_DISABLE_XET", raising = False)
+    monkeypatch.delenv("HF_HUB_ENABLE_HF_TRANSFER", raising = False)
+    monkeypatch.setattr(worker, "_activate_transformers_version_or_warn", fake_activate)
+    monkeypatch.setattr(hw, "detect_hardware", fake_detect_hardware)
+
+    event_queue = queue.Queue()
+    worker.run_mlx_training_process(
+        event_queue = event_queue,
+        stop_queue = queue.Queue(),
+        config = {"model_name": "mlx-community/Gemma-4-12B", "disable_xet": True},
+    )
+
+    event = event_queue.get_nowait()
+    assert order == [("activate", "mlx-community/Gemma-4-12B", None), "detect"]
+    assert os.environ["HF_HUB_DISABLE_XET"] == "1"
+    assert os.environ["HF_HUB_ENABLE_HF_TRANSFER"] == "0"
+    assert "MLX training requires Apple Silicon" in event["error"]
+
+
 if __name__ == "__main__":
     unittest.main()
diff --git a/studio/backend/tests/test_training_worker_flash_attn.py b/studio/backend/tests/test_training_worker_flash_attn.py
index 3c5d6cd094..7e7fc1af48 100644
--- a/studio/backend/tests/test_training_worker_flash_attn.py
+++ b/studio/backend/tests/test_training_worker_flash_attn.py
@@ -59,7 +59,6 @@ def test_runtime_flash_attn_prefers_prebuilt_wheel(monkeypatch):
     statuses: list[str] = []
 
     monkeypatch.delenv(worker._FLASH_ATTN_SKIP_ENV, raising = False)
-    monkeypatch.setattr(worker, "has_blackwell_gpu", lambda: False)
     monkeypatch.setattr(builtins, "__import__", _missing_flash_attn_import())
     monkeypatch.setattr(
         worker,
@@ -88,7 +87,6 @@ def test_runtime_flash_attn_falls_back_to_pypi(monkeypatch):
     statuses: list[str] = []
 
     monkeypatch.delenv(worker._FLASH_ATTN_SKIP_ENV, raising = False)
-    monkeypatch.setattr(worker, "has_blackwell_gpu", lambda: False)
     monkeypatch.setattr(builtins, "__import__", _missing_flash_attn_import())
     monkeypatch.setattr(
         worker,
@@ -141,27 +139,6 @@ def test_runtime_flash_attn_skip_env_avoids_all_install_work(monkeypatch):
     worker._sp.run.assert_not_called()
 
 
-def test_runtime_flash_attn_skips_on_blackwell(monkeypatch):
-    statuses: list[str] = []
-    install_mock = mock.Mock()
-
-    monkeypatch.delenv(worker._FLASH_ATTN_SKIP_ENV, raising = False)
-    monkeypatch.setattr(worker, "_should_try_runtime_flash_attn_install", lambda max_seq: True)
-    monkeypatch.setattr(worker, "has_blackwell_gpu", lambda: True)
-    monkeypatch.setattr(worker, "_install_package_wheel_first", install_mock)
-    monkeypatch.setattr(
-        worker,
-        "_send_status",
-        lambda queue, message: statuses.append(message),
-    )
-
-    worker._ensure_flash_attn_for_long_context(event_queue = [], max_seq_length = 65536)
-
-    install_mock.assert_not_called()
-    assert len(statuses) == 1
-    assert "Blackwell" in statuses[0]
-
-
 def test_causal_conv1d_fast_path_preserves_wheel_first_install_args(monkeypatch):
     install_mock = mock.Mock(return_value = True)
     monkeypatch.setattr(worker, "_install_package_wheel_first", install_mock)
diff --git a/studio/backend/tests/test_training_worker_import_discipline.py b/studio/backend/tests/test_training_worker_import_discipline.py
new file mode 100644
index 0000000000..a047c91704
--- /dev/null
+++ b/studio/backend/tests/test_training_worker_import_discipline.py
@@ -0,0 +1,81 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Invariant: the training worker must not import ``transformers`` before it activates the
+transformers sidecar.
+
+``core/training/worker.py:run_training_process`` runs a preflight (Xet decision, logging, hardware
+detection) and only THEN calls ``_activate_transformers_version`` -> ``activate_transformers_for_subprocess``,
+which prepends the correct ``.venv_t5_*`` (5.x) sidecar to ``sys.path``. Because activation only edits
+``sys.path``, it is a no-op for any module already cached in ``sys.modules``. So if the preflight imports
+``transformers`` (directly or transitively via ``unsloth_zoo``), the default 4.57.x gets pinned before
+the sidecar is on the path -- and 5.x models (Qwen3.5, GLM-4.7, gemma-4) then fail to load their
+tokenizer/config ("Tokenizer class TokenizersBackend does not exist").
+
+This regression shipped once when ``utils/hf_xet_fallback.py`` eagerly imported ``unsloth_zoo`` (which
+imports ``transformers``) at module load; the worker imports that shim during preflight to decide the
+Xet env flip (see issue #6951). This test locks the invariant in a fresh interpreter. It is CPU-only,
+needs no network/GPU/weights/sidecars, so it runs in the standard ``studio-backend-ci`` matrix.
+"""
+
+from __future__ import annotations
+
+import subprocess
+import sys
+from pathlib import Path
+
+_BACKEND_DIR = Path(__file__).resolve().parent.parent  # studio/backend
+
+# Mirrors run_training_process's imports that run BEFORE _activate_transformers_version (worker.py);
+# keep in sync. torch-dependent imports are optional (a no-torch CI shard skips them) but must still
+# not drag in transformers.
+_PREFLIGHT_SNIPPET = r"""
+import sys
+
+# worker.py: from utils.hf_xet_fallback import child_should_disable_xet  (+ call it)
+from utils.hf_xet_fallback import child_should_disable_xet
+child_should_disable_xet({})
+
+# worker.py: from loggers.config import LogConfig
+from loggers.config import LogConfig  # noqa: F401
+
+# worker.py: from utils.hardware import hardware  (imports torch, not transformers)
+try:
+    from utils.hardware import hardware as _hw  # noqa: F401
+except Exception:
+    pass  # torch may be absent in a no-torch shard; the invariant below still applies
+
+# worker.py: from .training import is_apple_silicon_training_platform, should_use_mlx_training_backend
+# (the MLX-dispatch preflight; must also stay clear of transformers). Guarded because it may pull
+# unsloth/trl, absent in a minimal shard -- but a partial import that leaked transformers would still
+# be caught by the assertion below.
+try:
+    from core.training.training import (  # noqa: F401
+        is_apple_silicon_training_platform as _is_apple,
+        should_use_mlx_training_backend as _use_mlx,
+    )
+except Exception:
+    pass
+
+leaked_tf = sorted(m for m in sys.modules if m == "transformers" or m.startswith("transformers."))
+leaked_zoo = sorted(m for m in sys.modules if m == "unsloth_zoo" or m.startswith("unsloth_zoo."))
+assert not leaked_tf, f"transformers imported during worker preflight (before sidecar activation): {leaked_tf}"
+assert not leaked_zoo, f"unsloth_zoo imported during worker preflight (before sidecar activation): {leaked_zoo}"
+print("PREFLIGHT_CLEAN")
+"""
+
+
+def test_worker_preflight_does_not_import_transformers():
+    """A fresh interpreter running the worker's pre-activation imports must leave ``transformers``
+    (and ``unsloth_zoo``) unimported, so the 5.x sidecar prepend is not defeated by a stale module."""
+    result = subprocess.run(
+        [sys.executable, "-c", _PREFLIGHT_SNIPPET],
+        cwd = str(_BACKEND_DIR),
+        capture_output = True,
+        text = True,
+    )
+    assert result.returncode == 0, (
+        "Worker preflight imported transformers before sidecar activation.\n"
+        f"STDOUT:\n{result.stdout}\nSTDERR:\n{result.stderr}"
+    )
+    assert "PREFLIGHT_CLEAN" in result.stdout, result.stdout
diff --git a/studio/backend/tests/test_worker_activates_correct_transformers.py b/studio/backend/tests/test_worker_activates_correct_transformers.py
new file mode 100644
index 0000000000..fe7b8dd25a
--- /dev/null
+++ b/studio/backend/tests/test_worker_activates_correct_transformers.py
@@ -0,0 +1,155 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Invariant: after the training worker runs its preflight and then activates the transformers
+sidecar, the in-process ``transformers`` must be the sidecar version the model requires -- not the
+default 4.57.x that the base environment ships.
+
+The CPU-only "does it choose the correct transformers version" guard, stronger than the pure
+import-order check in ``test_training_worker_import_discipline.py``: it runs the REAL tier detection
+(``get_transformers_tier``) and REAL activation (``activate_transformers_for_subprocess``) for a
+transformers-5.x model (Qwen3.5, tier 530) and asserts the version actually switched. It catches the
+whole failure family at once:
+
+  * a stale pre-activation ``transformers`` import (the #6951 / ``TokenizersBackend`` regression: an
+    already-cached 4.57.x defeats the sidecar's ``sys.path`` prepend),
+  * a wrong tier selected for a 5.x model, and
+  * activation not actually swapping the resident module.
+
+Why the CUDA spoof matters (verified): ``unsloth_zoo``'s eager ``import transformers`` only happens on
+its full, GPU-present init path. On a GPU-less runner it silently degrades and never preloads
+transformers -- which would MASK the stale-import bug (the check would falsely pass). Spoofing
+``torch.cuda`` so ``unsloth_zoo`` believes a GPU is present forces the real init path, exposing the
+regression on CPU CI. The spoof mirrors ``tests/_zoo_aggressive_cuda_spoof.py`` but is inlined so the
+test is self-contained in the ``studio-backend-ci`` matrix (whose conftest does not apply the shared
+spoof). No GPU/network/weights/real sidecar needed: a one-line stub sidecar stands in for the 5.x venv,
+so we only assert activation lands on it.
+
+Proven: passes on the fixed tree (active == 5.3.0) and fails on the buggy tree (active == 4.57.x) on
+a simulated GPU-less runner.
+"""
+
+from __future__ import annotations
+
+import subprocess
+import sys
+from pathlib import Path
+
+_BACKEND_DIR = Path(__file__).resolve().parent.parent  # studio/backend
+# Canonical CUDA spoof at the repo root (studio/backend -> studio -> repo root). Loaded by the
+# subprocess when present (matches the consolidated CI); absent in a standalone studio checkout, where
+# the subprocess falls back to a minimal inline spoof.
+_SPOOF_PATH = _BACKEND_DIR.parent.parent / "tests" / "_zoo_aggressive_cuda_spoof.py"
+
+# Runs in a fresh interpreter with cwd == studio/backend so ``utils.*`` resolves like the worker.
+# STUB_HOME (a pytest tmp dir) holds a throwaway ``.venv_t5_530`` sidecar exporting transformers 5.3.0.
+_SNIPPET = r"""
+import os, sys
+sys.path.insert(0, os.getcwd())
+
+# CUDA spoof so unsloth_zoo takes its full, transformers-importing init path on a GPU-less runner.
+# Without it unsloth_zoo degrades and never preloads transformers, which would MASK the stale-import
+# regression under test (verified). Prefer the repo's canonical spoof (single source of truth, and the
+# one the consolidated CI already relies on); fall back to a minimal inline spoof so this also works in
+# a standalone studio checkout. If torch is absent the fixed tree still passes below; the bug just
+# would not be exposable in that shard.
+try:
+    import torch  # noqa: F401
+    _sp = os.environ.get("SPOOF_PATH")
+    if _sp and os.path.exists(_sp):
+        import importlib.util
+        _spec = importlib.util.spec_from_file_location("_zoo_aggressive_cuda_spoof", _sp)
+        _mod = importlib.util.module_from_spec(_spec)
+        _spec.loader.exec_module(_mod)
+        _mod.apply()
+    else:
+        torch.cuda.is_available = lambda: True
+        torch.cuda.device_count = lambda: 1
+        torch.cuda.current_device = lambda: 0
+        torch.cuda.get_device_capability = lambda *a, **k: (8, 0)
+        torch.cuda.get_device_name = lambda *a, **k: "NVIDIA A100-SPOOFED"
+        torch.cuda.is_bf16_supported = lambda *a, **k: True
+        class _Props:
+            name = "NVIDIA A100-SPOOFED"
+            major = 8
+            minor = 0
+            total_memory = 80 * 1024**3
+            multi_processor_count = 108
+        torch.cuda.get_device_properties = lambda *a, **k: _Props()
+        torch.cuda.mem_get_info = lambda *a, **k: (0, 80 * 1024**3)
+except Exception:
+    pass
+os.environ["UNSLOTH_IS_PRESENT"] = "1"
+
+# Stub 5.x sidecar: activation only edits sys.path, so a package that merely exports __version__ is
+# enough to prove the resident transformers switched to it.
+home = os.environ["STUB_HOME"]
+pkg = os.path.join(home, ".venv_t5_530", "transformers")
+os.makedirs(pkg, exist_ok = True)
+with open(os.path.join(pkg, "__init__.py"), "w") as f:
+    f.write('__version__ = "5.3.0"\n')
+os.environ["UNSLOTH_STUDIO_HOME"] = home
+
+# Faithful worker preflight (worker.py: from utils.hf_xet_fallback import child_should_disable_xet).
+# This is the exact stale-import trigger: on the buggy tree it pulls unsloth_zoo -> transformers 4.57.x
+# into sys.modules BEFORE activation.
+from utils.hf_xet_fallback import child_should_disable_xet
+child_should_disable_xet({})
+_tf = sys.modules.get("transformers")
+preload = _tf.__version__ if _tf is not None else None
+
+# Real tier detection + real activation, with the 530 sidecar pointed at the stub above.
+import utils.transformers_version as tv
+tv._VENV_T5_530_DIR = os.path.join(home, ".venv_t5_530")
+tv._ensure_venv_t5_530_exists = lambda: True
+tier = tv.get_transformers_tier("Qwen/Qwen3.5-9B", None)
+tv.activate_transformers_for_subprocess("Qwen/Qwen3.5-9B", None)
+
+import transformers
+print(f"RESULT tier={tier} preload={preload} active={transformers.__version__}")
+"""
+
+
+def _parse(stdout: str) -> dict[str, str]:
+    for line in stdout.splitlines():
+        if line.startswith("RESULT "):
+            return dict(kv.split("=", 1) for kv in line.split()[1:])
+    return {}
+
+
+def test_worker_activates_correct_transformers_version(tmp_path):
+    """The worker's real preflight + activation for a transformers-5.x model (Qwen3.5, tier 530) must
+    leave the in-process ``transformers`` on the 5.x sidecar. A stale pre-activation import leaves the
+    default 4.57.x pinned and fails this assertion -- exactly the #6951 ``TokenizersBackend`` regression."""
+    result = subprocess.run(
+        [sys.executable, "-c", _SNIPPET],
+        cwd = str(_BACKEND_DIR),
+        env = {
+            **__import__("os").environ,
+            "STUB_HOME": str(tmp_path),
+            **({"SPOOF_PATH": str(_SPOOF_PATH)} if _SPOOF_PATH.exists() else {}),
+        },
+        capture_output = True,
+        text = True,
+    )
+    assert result.returncode == 0, (
+        "Worker preflight + activation harness crashed.\n"
+        f"STDOUT:\n{result.stdout}\nSTDERR:\n{result.stderr}"
+    )
+    parsed = _parse(result.stdout)
+    assert parsed, f"No RESULT line.\nSTDOUT:\n{result.stdout}\nSTDERR:\n{result.stderr}"
+
+    # Correct tier chosen for a transformers-5.x model (pure, deterministic; no network/GPU).
+    assert parsed["tier"] == "530", (
+        f"Wrong transformers tier for Qwen3.5 (expected 530, got {parsed['tier']}). "
+        "Tier detection regressed."
+    )
+
+    # Activation must actually swap the resident transformers to the sidecar version. If a preflight
+    # import cached 4.57.x first, the sidecar prepend is a no-op and this stays 4.57.x -- the bug.
+    assert parsed["active"] == "5.3.0", (
+        "Sidecar activation did NOT switch the in-process transformers to the model's 5.x version "
+        f"(active={parsed['active']}, preloaded-before-activation={parsed['preload']}). A pre-activation "
+        "transformers import (directly or via unsloth_zoo) defeated the sidecar; 5.x models (Qwen3.5, "
+        "GLM-4.7, gemma-4) then fail with 'Tokenizer class TokenizersBackend does not exist'. See #6951."
+    )
diff --git a/studio/backend/utils/coding_agents.py b/studio/backend/utils/coding_agents.py
new file mode 100644
index 0000000000..f7dd2f8357
--- /dev/null
+++ b/studio/backend/utils/coding_agents.py
@@ -0,0 +1,39 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Detect which `unsloth start <agent>` coding-agent CLIs are on PATH.
+
+The web UI only ever shows the user the "claude" flavor of the `unsloth start`
+command (see agent-command.ts), leaving anyone using Codex, OpenCode, and the
+other supported agents to manually edit the copied command. This module gives
+the frontend a way to ask which of those CLIs are actually installed so it can
+default to one the user can run immediately.
+"""
+
+import shutil
+
+# Keep in sync with the `unsloth start <agent>` subcommands defined in
+# unsloth_cli/commands/start.py. Each entry is the exact executable name that
+# subcommand launches, so a hit here means `unsloth start <agent>` can find the
+# binary on PATH without the user installing anything first.
+CODING_AGENTS: tuple[str, ...] = ("claude", "codex", "openclaw", "opencode", "hermes", "pi")
+
+
+def _is_on_path(agent: str) -> bool:
+    # shutil.which is documented to return None on a miss, but PATH lookups can
+    # still raise (e.g. a permission error while probing a directory entry);
+    # this is an advisory check, so a lookup failure should read as "not
+    # installed" instead of breaking the settings endpoint.
+    try:
+        return shutil.which(agent) is not None
+    except OSError:
+        return False
+
+
+def detect_installed_coding_agents() -> list[str]:
+    """Return the subset of CODING_AGENTS whose CLI binary is on PATH.
+
+    Order follows CODING_AGENTS, not discovery order, so callers can treat the
+    first entry as the preferred default among the installed agents.
+    """
+    return [agent for agent in CODING_AGENTS if _is_on_path(agent)]
diff --git a/studio/backend/utils/datasets/completion_masking.py b/studio/backend/utils/datasets/completion_masking.py
new file mode 100644
index 0000000000..c7c4a474e3
--- /dev/null
+++ b/studio/backend/utils/datasets/completion_masking.py
@@ -0,0 +1,144 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Completion-only masking policy shared by the CUDA and MLX training paths.
+
+Decides how train_on_responses_only is applied for a model: chat template
+auto-detection first, manual TEMPLATE_TO_RESPONSES_MAPPER markers as the
+fallback. gpt-oss included: its quantized checkpoints ship a different
+chat template, so only detection from the actual template is reliable.
+"""
+
+from .model_mappings import (
+    MODEL_TO_TEMPLATE_MAPPER,
+    TEMPLATE_TO_RESPONSES_MAPPER,
+    is_gpt_oss_model_name,
+)
+
+
+def lookup_manual_markers(model_name):
+    """Return (template_name, instruction_part, response_part) from the
+    manual template table, with None parts when the model or template is
+    not mapped."""
+    template = MODEL_TO_TEMPLATE_MAPPER.get((model_name or "").lower())
+    markers = TEMPLATE_TO_RESPONSES_MAPPER.get(template) if template else None
+    if markers:
+        return template, markers["instruction"], markers["response"]
+    return template, None, None
+
+
+def apply_completion_masking(
+    trainer,
+    model_name,
+    train_fn,
+    num_proc = None,
+    notify = None,
+    detect_fn = None,
+):
+    """Apply completion-only masking with auto-detection first and the manual
+    template table as fallback.
+
+    Args:
+        trainer: The platform trainer (SFTTrainer or MLXTrainer).
+        model_name: Model repo id used for table lookup and the gpt-oss
+            renamed-checkpoint fallback.
+        train_fn: The platform train_on_responses_only callable.
+        num_proc: Forwarded to train_fn when not None (CUDA path only).
+        notify: Optional callback notify(level, message) with level "info" or
+            "warning" for user-visible progress and warnings.
+        detect_fn: Marker detector (tokenizer/processor) -> (instruction_part,
+            response_part). Defaults to unsloth_zoo's get_chat_template_parts,
+            which raises loudly when the template cannot be parsed. Test seam.
+
+    Returns:
+        (trainer, applied): the possibly wrapped trainer and whether masking
+        was applied. When applied is False the trainer is unchanged and
+        training runs on full sequences.
+
+    Only marker DETECTION failures trigger the table fallback. Exceptions
+    raised while applying the masking (dataset map, tokenization) propagate
+    to the caller in both the auto and manual paths, so a real failure stops
+    the run instead of silently changing the training objective.
+    """
+    if notify is None:
+        notify = lambda level, message: None
+    kwargs = {}
+    if num_proc is not None:
+        kwargs["num_proc"] = num_proc
+
+    template, instruction_part, response_part = lookup_manual_markers(model_name)
+
+    # gpt-oss goes auto-first: quantized/BF16 checkpoints ship a channel-less
+    # template, so the manual markers match nothing (zero tokens trained). Auto
+    # derives markers from whichever template ships, and per the harmony format
+    # only the final terminator carries stop supervision. Renamed checkpoints
+    # miss the exact-name table, so give the fallback the gpt-oss markers.
+    if is_gpt_oss_model_name(model_name) and not (instruction_part and response_part):
+        markers = TEMPLATE_TO_RESPONSES_MAPPER.get("gpt-oss")
+        if markers:
+            template = "gpt-oss"
+            instruction_part = markers["instruction"]
+            response_part = markers["response"]
+    processor = getattr(trainer, "processing_class", None) or getattr(trainer, "tokenizer", None)
+    # mlx-lm TokenizerWrapper hides underscore attrs, so preset _unsloth_*
+    # markers are invisible through it. Unwrap to the real tokenizer (as
+    # zoo's MLX resolver does) before the preset check and detection.
+    if type(processor).__name__ == "TokenizerWrapper":
+        wrapped = getattr(processor, "_tokenizer", None)
+        if wrapped is not None:
+            processor = wrapped
+    inner = getattr(processor, "tokenizer", processor)
+    if hasattr(inner, "_unsloth_input_part") and hasattr(inner, "_unsloth_output_part"):
+        # Markers preset on the tokenizer; zoo reuses them on a bare call.
+        trainer = train_fn(trainer, **kwargs)
+        notify(
+            "info",
+            "Train on responses only configured via tokenizer preset markers",
+        )
+        return trainer, True
+    auto_instruction = auto_response = None
+    try:
+        if detect_fn is None:
+            # Torch-backed import is fine: the MLX train_fn itself requires
+            # unsloth_zoo.dataset_utils, so a torch-free host cannot mask either way.
+            from unsloth_zoo.dataset_utils import get_chat_template_parts as detect_fn
+        auto_instruction, auto_response = detect_fn(processor)
+    except Exception as e:
+        notify(
+            "warning",
+            f"Auto-detection of instruction/response markers failed ({e}); "
+            f"falling back to the template table",
+        )
+    if auto_instruction and auto_response:
+        trainer = train_fn(
+            trainer,
+            instruction_part = auto_instruction,
+            response_part = auto_response,
+            **kwargs,
+        )
+        notify(
+            "info",
+            "Train on responses only configured via chat template auto-detection",
+        )
+        return trainer, True
+
+    if instruction_part and response_part:
+        trainer = train_fn(
+            trainer,
+            instruction_part = instruction_part,
+            response_part = response_part,
+            **kwargs,
+        )
+        notify(
+            "info",
+            f"Train on responses only configured with template table markers ({template})",
+        )
+        return trainer, True
+
+    notify(
+        "warning",
+        f"'Train on completions' could not be applied for {model_name}: no "
+        f"auto-detected or mapped instruction/response markers. Training "
+        f"will run on full sequences (prompts included).",
+    )
+    return trainer, False
diff --git a/studio/backend/utils/datasets/model_mappings.py b/studio/backend/utils/datasets/model_mappings.py
index 9d2c983aed..65ba4b4688 100644
--- a/studio/backend/utils/datasets/model_mappings.py
+++ b/studio/backend/utils/datasets/model_mappings.py
@@ -485,9 +485,11 @@ TEMPLATE_TO_RESPONSES_MAPPER = {
         "instruction": "<|im_start|>user\n",
         "response": "<|im_start|>assistant\n",
     },
+    # No "<think>" suffix: Qwen3-Thinking-2507 strips it from non-final turns
+    # and QwQ renders none, so a marker holding it masks those responses.
     "qwen3-thinking": {
         "instruction": "<|im_start|>user\n",
-        "response": "<|im_start|>assistant\n<think>",
+        "response": "<|im_start|>assistant\n",
     },
     "qwen3": {
         "instruction": "<|im_start|>user\n",
@@ -525,29 +527,39 @@ TEMPLATE_TO_RESPONSES_MAPPER = {
         "instruction": "<|im_start|>user<|im_sep|>",
         "response": "<|im_start|>assistant<|im_sep|>",
     },
+    # No surrounding spaces: in Mistral v0.3 they fold into neighbouring text
+    # tokens ("[INST]"/"[/INST]" are single special tokens), so padded strings
+    # never match and everything masks. Same for Llama-2's SentencePiece.
     "mistral": {
-        "instruction": "[INST] ",
-        "response": " [/INST]",
+        "instruction": "[INST]",
+        "response": "[/INST]",
     },
     "llama": {
-        "instruction": "[INST] ",
-        "response": " [/INST]",
+        # <s>-anchored: llama-2 tokenizes [INST] after <s> as bare "[" on
+        # transformers 5.x (standalone gives space-prefixed "▁["), so an
+        # unanchored marker misses every turn boundary there.
+        "instruction": "<s>[INST]",
+        "response": "[/INST]",
     },
     "chatml": {
         "instruction": "<|im_start|>user\n",
         "response": "<|im_start|>assistant\n",
     },
+    # Leading newline required: Zephyr's role tags are plain text, and
+    # SentencePiece tokenizes "<|assistant|>" differently at text start than
+    # after "</s>\n". Without the "\n" anchor the markers never match real
+    # turns, so every assistant token masks.
     "zephyr": {
-        "instruction": "<|user|>\n",
-        "response": "<|assistant|>\n",
+        "instruction": "\n<|user|>\n",
+        "response": "\n<|assistant|>\n",
     },
     "unsloth": {
-        "instruction": ">>> User: ",
-        "response": ">>> Assistant: ",
+        "instruction": ">>> User:",
+        "response": ">>> Assistant:",
     },
     "vicuna": {
-        "instruction": "USER: ",
-        "response": "ASSISTANT: ",
+        "instruction": "USER:",
+        "response": "ASSISTANT:",
     },
     "alpaca": {
         "instruction": "### Instruction:\n",
@@ -573,16 +585,21 @@ TEMPLATE_TO_RESPONSES_MAPPER = {
         "instruction": "<|im_start|>user\n",
         "response": "<|im_start|>assistant\n",
     },
+    # No trailing space: SentencePiece folds it into the next content token
+    # ("▁Hello"), so the padded marker never matches and masks everything.
     "starling": {
-        "instruction": "GPT4 Correct User: ",
-        "response": "GPT4 Correct Assistant: ",
+        "instruction": "GPT4 Correct User:",
+        "response": "GPT4 Correct Assistant:",
     },
     "yi-chat": {
         "instruction": "<|im_start|>user\n",
         "response": "<|im_start|>assistant\n",
     },
+    # "[gMASK]<sop>" appears once at text start, so a marker holding it matches
+    # no later user turn; "<think>" is scaffolding GLM-4.x renders as a lone
+    # "</think>" on non-final turns, so "<|assistant|><think>" never matches.
     "glm": {
-        "instruction": "[gMASK]<sop><|user|>",
-        "response": "<|assistant|><think>",
+        "instruction": "<|user|>",
+        "response": "<|assistant|>",
     },
 }
diff --git a/studio/backend/utils/hf_xet_fallback.py b/studio/backend/utils/hf_xet_fallback.py
index 2dd2247396..9bc4a60fad 100644
--- a/studio/backend/utils/hf_xet_fallback.py
+++ b/studio/backend/utils/hf_xet_fallback.py
@@ -6,6 +6,16 @@
 Re-exports the shared API and injects Studio's marker-aware cache purge
 (``prepare_cache_for_transport``) so the download manager keeps its ``.transport``
 marker semantics on the HTTP retry.
+
+Import discipline: ``unsloth_zoo``'s ``__init__`` eagerly imports ``transformers``. The workers
+import this shim at startup (to decide the per-worker Xet env flip) *before* activating the model's
+``transformers`` sidecar. Activation only prepends the sidecar to ``sys.path``, so a ``transformers``
+already cached in ``sys.modules`` (via an eager ``unsloth_zoo`` import here) wins -- pinning the
+default 4.57.x and regressing Qwen3.5 / GLM-4.7 / gemma-4 training with
+``Tokenizer class TokenizersBackend does not exist``. So the shared backend is loaded **lazily**
+(``_load_shared``), only on first use of a heavy download helper, i.e. after the sidecar is active.
+``child_should_disable_xet`` and the ``DEFAULT_*`` constants are defined locally so importing them
+never triggers the heavy load.
 """
 
 from __future__ import annotations
@@ -13,161 +23,230 @@ from __future__ import annotations
 import threading
 from typing import Any, Callable, Optional
 
-_shared_import_error = None
-try:
-    import unsloth_zoo.hf_xet_fallback as _shared
-    _shared_available = True
-except Exception as _exc:  # noqa: BLE001 - any import failure must degrade, not crash
-    # unsloth_zoo's __init__ runs torch/GPU detection, which raises on a torch-less/GPU-less Studio
-    # host. The download helper needs none of it, so retry via the light UNSLOTH_ZOO_DISABLE_GPU_INIT
-    # path before giving up.
-    _shared_import_error = _exc
-    import os as _os
+# Defaults mirror unsloth_zoo.hf_xet_fallback; plain literals so they resolve (including as
+# default args below) without importing unsloth_zoo/transformers.
+DEFAULT_GRACE_PERIOD = 10.0
+DEFAULT_HEARTBEAT_INTERVAL = 30.0
+DEFAULT_STALL_TIMEOUT = 180.0
 
-    _prev_gpu_init = _os.environ.get("UNSLOTH_ZOO_DISABLE_GPU_INIT")
-    _os.environ["UNSLOTH_ZOO_DISABLE_GPU_INIT"] = "1"
-    try:
-        import unsloth_zoo.hf_xet_fallback as _shared
-        _shared_available = True
-        _shared_import_error = None
-    except Exception as _exc2:  # noqa: BLE001 - degrade so Studio still boots with plain HF downloads
-        _shared_import_error = _exc2
-        _shared_available = False
-    finally:
-        if _prev_gpu_init is None:
-            _os.environ.pop("UNSLOTH_ZOO_DISABLE_GPU_INIT", None)
-        else:
-            _os.environ["UNSLOTH_ZOO_DISABLE_GPU_INIT"] = _prev_gpu_init
+# --- lazy shared-backend loader ----------------------------------------------------------------
+_shared: Any = None
+_shared_available: Optional[bool] = None  # None = not yet attempted
+_shared_import_error: Optional[BaseException] = None
+_load_lock = threading.Lock()
 
-if _shared_available:
-    # Bind by assignment so each public name shares one module-level binding with the degraded branch.
-    DEFAULT_GRACE_PERIOD = _shared.DEFAULT_GRACE_PERIOD
-    DEFAULT_HEARTBEAT_INTERVAL = _shared.DEFAULT_HEARTBEAT_INTERVAL
-    DEFAULT_STALL_TIMEOUT = _shared.DEFAULT_STALL_TIMEOUT
-    DownloadStallError = _shared.DownloadStallError
-    child_should_disable_xet = _shared.child_should_disable_xet
-    get_hf_download_state = _shared.get_hf_download_state
-    start_watchdog = _shared.start_watchdog
-    _shared_hf_hub_download_with_xet_fallback = _shared.hf_hub_download_with_xet_fallback
-    _shared_snapshot_download_with_xet_fallback = _shared.snapshot_download_with_xet_fallback
-else:
-    # Degrade instead of crashing Studio: plain HF downloads, stall watchdog disabled. Thin stubs,
-    # not a second copy of the orchestration; recovery returns once unsloth_zoo is upgraded.
-    import logging as _logging
 
-    _logging.getLogger(__name__).warning(
-        "unsloth_zoo.hf_xet_fallback unavailable (%s); the Xet stall watchdog is "
-        "disabled. Install/upgrade unsloth_zoo (and its torch dependency) to "
-        "re-enable automatic Xet -> HTTP download recovery.",
-        _shared_import_error,
-    )
+def _load_shared() -> bool:
+    """Import ``unsloth_zoo.hf_xet_fallback`` on demand; return True if available. Deferred so
+    importing this module at worker startup does not pull transformers in before the sidecar is
+    activated. Degrades (returns False) rather than crashing when unsloth_zoo is unavailable."""
+    global _shared, _shared_available, _shared_import_error
+    if _shared_available is not None:
+        return _shared_available
+    with _load_lock:
+        if _shared_available is not None:
+            return _shared_available
+        try:
+            import unsloth_zoo.hf_xet_fallback as shared
 
-    DEFAULT_HEARTBEAT_INTERVAL = 30.0
-    DEFAULT_STALL_TIMEOUT = 180.0
-    DEFAULT_GRACE_PERIOD = 10.0
+            _shared = shared
+            _shared_available = True
+            _shared_import_error = None
+            return True
+        except Exception as exc:  # noqa: BLE001 - any import failure must degrade, not crash
+            # unsloth_zoo's __init__ runs torch/GPU detection, which raises on a torch-less/GPU-less
+            # host. The download helper needs none of it, so retry via UNSLOTH_ZOO_DISABLE_GPU_INIT.
+            _shared_import_error = exc
+            import os as _os
 
-    class DownloadStallError(RuntimeError):
-        """Stub mirror so callers' ``except`` clauses resolve; never raised in degraded mode."""
+            _prev_gpu_init = _os.environ.get("UNSLOTH_ZOO_DISABLE_GPU_INIT")
+            _os.environ["UNSLOTH_ZOO_DISABLE_GPU_INIT"] = "1"
+            try:
+                import unsloth_zoo.hf_xet_fallback as shared
 
-    def child_should_disable_xet(config: dict) -> bool:
-        return bool(config.get("disable_xet"))
+                _shared = shared
+                _shared_available = True
+                _shared_import_error = None
+                return True
+            except Exception as exc2:  # noqa: BLE001 - degrade so Studio still boots with plain HF
+                _shared_import_error = exc2
+                _shared_available = False
+                import logging as _logging
 
-    def get_hf_download_state(*args: Any, **kwargs: Any) -> None:
-        return None  # unmeasurable -> the (absent) watchdog never fires
+                _logging.getLogger(__name__).warning(
+                    "unsloth_zoo.hf_xet_fallback unavailable (%s); the Xet stall watchdog is "
+                    "disabled. Install/upgrade unsloth_zoo (and its torch dependency) to "
+                    "re-enable automatic Xet -> HTTP download recovery.",
+                    _shared_import_error,
+                )
+                return False
+            finally:
+                if _prev_gpu_init is None:
+                    _os.environ.pop("UNSLOTH_ZOO_DISABLE_GPU_INIT", None)
+                else:
+                    _os.environ["UNSLOTH_ZOO_DISABLE_GPU_INIT"] = _prev_gpu_init
 
-    def start_watchdog(
-        *,
-        on_heartbeat: "Optional[Callable[[str], None]]" = None,
-        interval: float = DEFAULT_HEARTBEAT_INTERVAL,
-        xet_disabled: bool = False,
-        **kwargs: Any,
-    ) -> "threading.Event":
-        # No stall detection, but keep emitting heartbeats so the orchestrator's inactivity deadline
-        # is not tripped during a long download.
-        stop = threading.Event()
-        if on_heartbeat is None:
-            return stop
-        transport = "https" if xet_disabled else "xet"
 
-        def _beat() -> None:
-            while not stop.wait(interval):
-                try:
-                    on_heartbeat(f"Downloading ({transport} transport)...")
-                except Exception:
-                    pass
+def child_should_disable_xet(config: dict) -> bool:
+    """Single source of truth for the per-worker Xet env flip (mirrors
+    ``unsloth_zoo.hf_xet_fallback.child_should_disable_xet``). Deliberately lightweight: importing or
+    calling it must NOT pull in unsloth_zoo/transformers, so the worker can decide before activating
+    the transformers sidecar (see the module docstring)."""
+    return bool(config.get("disable_xet"))
 
-        threading.Thread(
-            target = _beat,
-            daemon = True,
-            name = "hf-xet-degraded-heartbeat",
-        ).start()
+
+# --- degraded stubs (used only when unsloth_zoo is unavailable) -------------------------------
+class _DegradedDownloadStallError(RuntimeError):
+    """Stub mirror so callers' ``except`` clauses resolve; never raised in degraded mode."""
+
+
+def _degraded_get_hf_download_state(*args: Any, **kwargs: Any) -> None:
+    return None  # unmeasurable -> the (absent) watchdog never fires
+
+
+def _degraded_start_watchdog(
+    *,
+    on_heartbeat: "Optional[Callable[[str], None]]" = None,
+    interval: float = DEFAULT_HEARTBEAT_INTERVAL,
+    xet_disabled: bool = False,
+    **kwargs: Any,
+) -> "threading.Event":
+    # No stall detection, but keep emitting heartbeats so the orchestrator's inactivity deadline
+    # is not tripped during a long download.
+    stop = threading.Event()
+    if on_heartbeat is None:
         return stop
+    transport = "https" if xet_disabled else "xet"
 
-    def _degraded_cancelled(cancel_event: "Optional[threading.Event]") -> bool:
-        return cancel_event is not None and cancel_event.is_set()
+    def _beat() -> None:
+        while not stop.wait(interval):
+            try:
+                on_heartbeat(f"Downloading ({transport} transport)...")
+            except Exception:
+                pass
 
-    def _shared_hf_hub_download_with_xet_fallback(
-        repo_id: str,
-        filename: str,
-        token: Optional[str],
-        *,
-        repo_type: str = "model",
-        revision: Optional[str] = None,
-        cache_dir: Optional[str] = None,
-        force_download: bool = False,
-        cancel_event: "Optional[threading.Event]" = None,
-        **_ignored: Any,
-    ) -> str:
-        # Keep the cancellation contract: do not start or return a download once cancelled.
-        if _degraded_cancelled(cancel_event):
-            raise RuntimeError("Cancelled")
+    threading.Thread(
+        target = _beat,
+        daemon = True,
+        name = "hf-xet-degraded-heartbeat",
+    ).start()
+    return stop
 
-        from huggingface_hub import hf_hub_download
 
-        path = hf_hub_download(
-            repo_id = repo_id,
-            filename = filename,
-            token = token,
-            repo_type = repo_type,
-            revision = revision,
-            cache_dir = cache_dir,
-            force_download = force_download,
-        )
-        if _degraded_cancelled(cancel_event):
-            raise RuntimeError("Cancelled")
-        return path
+def _degraded_cancelled(cancel_event: "Optional[threading.Event]") -> bool:
+    return cancel_event is not None and cancel_event.is_set()
 
-    def _shared_snapshot_download_with_xet_fallback(
-        repo_id: str,
-        *,
-        revision: Optional[str] = None,
-        token: Optional[str] = None,
-        repo_type: str = "model",
-        cache_dir: Optional[str] = None,
-        allow_patterns: Optional[Any] = None,
-        ignore_patterns: Optional[Any] = None,
-        force_download: bool = False,
-        cancel_event: "Optional[threading.Event]" = None,
-        **_ignored: Any,
-    ) -> str:
-        if _degraded_cancelled(cancel_event):
-            raise RuntimeError("Cancelled")
 
-        from huggingface_hub import snapshot_download
+def _degraded_hf_hub_download_with_xet_fallback(
+    repo_id: str,
+    filename: str,
+    token: Optional[str],
+    *,
+    repo_type: str = "model",
+    revision: Optional[str] = None,
+    cache_dir: Optional[str] = None,
+    force_download: bool = False,
+    cancel_event: "Optional[threading.Event]" = None,
+    **_ignored: Any,
+) -> str:
+    # Keep the cancellation contract: do not start or return a download once cancelled.
+    if _degraded_cancelled(cancel_event):
+        raise RuntimeError("Cancelled")
 
-        path = snapshot_download(
-            repo_id = repo_id,
-            repo_type = repo_type,
-            revision = revision,
-            token = token,
-            cache_dir = cache_dir,
-            allow_patterns = allow_patterns,
-            ignore_patterns = ignore_patterns,
-            force_download = force_download,
-        )
-        if _degraded_cancelled(cancel_event):
-            raise RuntimeError("Cancelled")
-        return path
+    from huggingface_hub import hf_hub_download
+
+    path = hf_hub_download(
+        repo_id = repo_id,
+        filename = filename,
+        token = token,
+        repo_type = repo_type,
+        revision = revision,
+        cache_dir = cache_dir,
+        force_download = force_download,
+    )
+    if _degraded_cancelled(cancel_event):
+        raise RuntimeError("Cancelled")
+    return path
+
+
+def _degraded_snapshot_download_with_xet_fallback(
+    repo_id: str,
+    *,
+    revision: Optional[str] = None,
+    token: Optional[str] = None,
+    repo_type: str = "model",
+    cache_dir: Optional[str] = None,
+    allow_patterns: Optional[Any] = None,
+    ignore_patterns: Optional[Any] = None,
+    force_download: bool = False,
+    cancel_event: "Optional[threading.Event]" = None,
+    **_ignored: Any,
+) -> str:
+    if _degraded_cancelled(cancel_event):
+        raise RuntimeError("Cancelled")
+
+    from huggingface_hub import snapshot_download
+
+    path = snapshot_download(
+        repo_id = repo_id,
+        repo_type = repo_type,
+        revision = revision,
+        token = token,
+        cache_dir = cache_dir,
+        allow_patterns = allow_patterns,
+        ignore_patterns = ignore_patterns,
+        force_download = force_download,
+    )
+    if _degraded_cancelled(cancel_event):
+        raise RuntimeError("Cancelled")
+    return path
+
+
+# --- lazy attribute access for the heavy shared API -------------------------------------------
+# ``DownloadStallError`` (class identity matters for ``except``), ``start_watchdog`` and
+# ``get_hf_download_state`` come from the shared backend when available, else the degraded stubs.
+# Resolved via PEP 562 ``__getattr__`` so ``from utils.hf_xet_fallback import X`` triggers the load
+# only for these heavy names, not for ``child_should_disable_xet`` / ``DEFAULT_*``.
+_DEGRADED_ATTRS = {
+    "DownloadStallError": _DegradedDownloadStallError,
+    "start_watchdog": _degraded_start_watchdog,
+    "get_hf_download_state": _degraded_get_hf_download_state,
+}
+
+# Annotation-only declarations for the three names above: they bind NO value, so lookup still misses
+# and PEP 562 ``__getattr__`` resolves them lazily -- but ruff/pyflakes see them as defined, so listing
+# them in ``__all__`` does not trip F822 (while F822 still catches a real typo elsewhere in the list).
+DownloadStallError: type
+start_watchdog: Any
+get_hf_download_state: Any
+
+
+def __getattr__(name: str) -> Any:
+    if name in _DEGRADED_ATTRS:
+        if _load_shared():
+            return getattr(_shared, name)
+        return _DEGRADED_ATTRS[name]
+    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
+
+
+# Indirection seam the public wrappers call (and tests monkeypatch): lazy-load the shared backend,
+# then dispatch to it or the degraded stub. The ``_shared_*`` names preserve the pre-refactor contract.
+def _shared_hf_hub_download_with_xet_fallback(*args: Any, **kwargs: Any) -> str:
+    impl = (
+        _shared.hf_hub_download_with_xet_fallback
+        if _load_shared()
+        else _degraded_hf_hub_download_with_xet_fallback
+    )
+    return impl(*args, **kwargs)
+
+
+def _shared_snapshot_download_with_xet_fallback(*args: Any, **kwargs: Any) -> str:
+    impl = (
+        _shared.snapshot_download_with_xet_fallback
+        if _load_shared()
+        else _degraded_snapshot_download_with_xet_fallback
+    )
+    return impl(*args, **kwargs)
 
 
 __all__ = [
diff --git a/studio/backend/utils/llama_cpp_update.py b/studio/backend/utils/llama_cpp_update.py
index c16ae91467..1bcbfbf95a 100644
--- a/studio/backend/utils/llama_cpp_update.py
+++ b/studio/backend/utils/llama_cpp_update.py
@@ -514,6 +514,12 @@ def _run_update(install_dir: Path, repo: str, asset: Optional[str], script: Path
         logger.info("llama update: installing", cmd = " ".join(cmd))
         # Stream progress lines into job["progress"].
         env = dict(os.environ, UNSLOTH_PROGRESS_PERCENT_STEP = "5")
+        # Preserve a Vulkan install across updates: detect_host on a CUDA/ROCm
+        # box would otherwise re-route and silently replace the Vulkan build.
+        # Re-assert it via the same env flag setup uses (mirrors
+        # _rocm_install_args).
+        if asset and "vulkan" in asset.lower():
+            env["UNSLOTH_FORCE_VULKAN"] = "1"
         proc = subprocess.Popen(
             cmd,
             stdout = subprocess.PIPE,
diff --git a/studio/backend/utils/models/model_config.py b/studio/backend/utils/models/model_config.py
index 5d8458e5f0..281ca24281 100644
--- a/studio/backend/utils/models/model_config.py
+++ b/studio/backend/utils/models/model_config.py
@@ -1617,36 +1617,60 @@ def _iter_hf_cache_snapshots(repo_id: str):
 
     cache_dir = Path(hf_constants.HF_HUB_CACHE)
     target = f"models--{repo_id.replace('/', '--')}".lower()
-    repo_dir: Optional[Path] = None
+    repo_dirs: list[Path] = []
     try:
         if not cache_dir.is_dir():
             return
         for entry in cache_dir.iterdir():
             if entry.is_dir() and entry.name.lower() == target:
-                repo_dir = entry
-                break
+                repo_dirs.append(entry)
     except OSError:
         return
-    if repo_dir is None:
+    if not repo_dirs:
         return
 
-    snapshots = repo_dir / "snapshots"
-    try:
-        if not snapshots.is_dir():
-            return
-        snap_dirs = [s for s in snapshots.iterdir() if s.is_dir()]
-    except OSError:
+    snap_dirs: list[Path] = []
+    for repo_dir in repo_dirs:
+        snapshots = repo_dir / "snapshots"
+        try:
+            if snapshots.is_dir():
+                for snap_dir in snapshots.iterdir():
+                    try:
+                        if snap_dir.is_dir():
+                            snap_dirs.append(snap_dir)
+                    except OSError:
+                        continue
+        except OSError:
+            continue
+    if not snap_dirs:
         return
-    snap_dirs.sort(key = lambda s: s.stat().st_mtime, reverse = True)
-    yield from snap_dirs
+    snap_dirs_with_mtime = []
+    for snap_dir in snap_dirs:
+        try:
+            snap_dirs_with_mtime.append((snap_dir.stat().st_mtime, snap_dir))
+        except OSError:
+            continue
+    snap_dirs_with_mtime.sort(key = lambda item: item[0], reverse = True)
+    yield from (snap_dir for _, snap_dir in snap_dirs_with_mtime)
 
 
 def _list_gguf_variants_from_hf_cache(repo_id: str) -> Optional[tuple[list[GgufVariantInfo], bool]]:
-    """Variants from the local HF cache snapshot, or None if not cached."""
+    """Variants from the local HF cache snapshot, or None if not cached.
+
+    A newer snapshot can hold only a companion file (for example a vision
+    projector fetched on demand) while the quant files live in an older
+    snapshot. Returning the first snapshot that merely reports a vision flag
+    would shadow those real variants, so keep scanning older snapshots for
+    actual variants and carry the vision flag across snapshots.
+    """
+    any_vision = False
     for snap in _iter_hf_cache_snapshots(repo_id):
         variants, has_vision = list_local_gguf_variants(str(snap))
-        if variants or has_vision:
-            return variants, has_vision
+        any_vision = any_vision or has_vision
+        if variants:
+            return variants, any_vision
+    if any_vision:
+        return [], True
     return None
 
 
diff --git a/studio/backend/utils/transformers_version.py b/studio/backend/utils/transformers_version.py
index 2a63caa5b2..a69673f081 100644
--- a/studio/backend/utils/transformers_version.py
+++ b/studio/backend/utils/transformers_version.py
@@ -28,7 +28,9 @@ Strategy:
   sys.path swap using the same directories pre-installed by setup.sh.
 """
 
+import ast
 import importlib
+import importlib.util
 import json
 import structlog
 from loggers import get_logger
@@ -173,6 +175,7 @@ _TRANSFORMERS_530_ARCHITECTURES: set[str] = {
     "Qwen3MoeForCausalLM",
     "Qwen3NextForCausalLM",
     "Glm4MoeLiteForCausalLM",
+    "Lfm2MoeForCausalLM",
     "Lfm2VlForConditionalGeneration",
 }
 _TRANSFORMERS_530_MODEL_TYPES: set[str] = {
@@ -183,6 +186,7 @@ _TRANSFORMERS_530_MODEL_TYPES: set[str] = {
     "qwen3_moe",
     "qwen3_next",
     "glm4_moe_lite",
+    "lfm2_moe",
     "lfm2_vl",
 }
 
@@ -870,6 +874,116 @@ def _cached_config_json(model_name: str, hf_token: str | None) -> dict | None:
     return _config_json_cache.get(_token_cache_key(model_name, hf_token))
 
 
+# --- Static tier from CONFIG_MAPPING_NAMES (AST only: no import/network/exec) ---
+# A model_type absent from an overlay's mapping can't load there. Parse each sidecar's
+# config map from source and pick the lowest tier that ships it, so a new arch routes
+# correctly with no per-model table edit. Only ever upgrades default, never lowers.
+_config_mapping_cache: dict[str, frozenset[str]] = {}
+
+
+def _overlay_transformers_dir(tier: str) -> str | None:
+    """transformers source dir for a tier, located without importing it."""
+    if tier != "default":
+        root = {"530": _VENV_T5_530_DIR, "550": _VENV_T5_550_DIR, "510": _VENV_T5_510_DIR}.get(tier)
+        src = os.path.join(root, "transformers") if root else None
+        return src if src and _safe_is_dir(Path(src)) else None
+    # default: the base 4.x transformers. find_spec resolves to a 5.x sidecar if one
+    # is already on sys.path, so skip any .venv_t5_* / llmcompressor overlay dir.
+    sidecars = tuple(
+        os.path.abspath(d) + os.sep
+        for d in (_VENV_T5_530_DIR, _VENV_T5_550_DIR, _VENV_T5_510_DIR, _VENV_LLMCOMPRESSOR_DIR)
+    )
+    candidates = []
+    try:
+        spec = importlib.util.find_spec("transformers")
+        if spec and spec.origin:
+            candidates.append(os.path.dirname(spec.origin))
+    except Exception:
+        pass
+    candidates += [os.path.join(e, "transformers") for e in sys.path if e]
+    for c in candidates:
+        if _safe_is_dir(Path(c)) and not os.path.abspath(c).startswith(sidecars):
+            return c
+    return None
+
+
+def _mapping_first_keys(value: ast.AST) -> set[str]:
+    """First keys of a dict literal, or of an OrderedDict(...)/dict(...)/.update(...)
+    built from 2-tuple lists and **{...} unpacking."""
+
+    def keys_of(node):
+        if isinstance(node, ast.Dict):
+            return list(node.keys)
+        if isinstance(node, (ast.List, ast.Tuple)):
+            return [
+                el.elts[0] for el in node.elts if isinstance(el, (ast.Tuple, ast.List)) and el.elts
+            ]
+        return []
+
+    nodes = keys_of(value)
+    if isinstance(value, ast.Call):
+        for a in value.args:
+            nodes += keys_of(a)
+        for kw in value.keywords:  # **{...} unpacking has kw.arg is None
+            if kw.arg is None:
+                nodes += keys_of(kw.value)
+    return {n.value for n in nodes if isinstance(n, ast.Constant) and isinstance(n.value, str)}
+
+
+def _config_model_types(tier: str) -> frozenset[str]:
+    """model_type keys in a tier's CONFIG_MAPPING_NAMES (5.10 moved it to auto_mappings.py)."""
+    cached = _config_mapping_cache.get(tier)
+    if cached is not None:
+        return cached
+    tdir = _overlay_transformers_dir(tier)
+    if tdir is None:
+        return frozenset()  # overlay not provisioned yet; do not cache so a later call re-reads
+    keys: set[str] = set()
+    for rel in ("models/auto/configuration_auto.py", "models/auto/auto_mappings.py"):
+        path = Path(tdir) / rel
+        if not _safe_is_file(path):
+            continue
+        try:
+            tree = ast.parse(path.read_text(encoding = "utf-8"))
+            for node in ast.walk(tree):
+                # direct binding, or a CONFIG_MAPPING_NAMES.update({...}) mutation
+                if isinstance(node, ast.Assign) and any(
+                    isinstance(t, ast.Name) and t.id == "CONFIG_MAPPING_NAMES" for t in node.targets
+                ):
+                    keys |= _mapping_first_keys(node.value)
+                elif isinstance(node, ast.Expr) and isinstance(node.value, ast.Call):
+                    fn = node.value.func
+                    if (
+                        isinstance(fn, ast.Attribute)
+                        and fn.attr == "update"
+                        and isinstance(fn.value, ast.Name)
+                        and fn.value.id == "CONFIG_MAPPING_NAMES"
+                    ):
+                        keys |= _mapping_first_keys(node.value)
+        except Exception:
+            continue
+    result = frozenset(keys)
+    _config_mapping_cache[tier] = result
+    return result
+
+
+def _tier_from_config_mapping(cfg: dict) -> str | None:
+    """Lowest tier whose transformers ships cfg's model_type, or None if unknown."""
+    model_type = cfg.get("model_type")
+    if not isinstance(model_type, str):
+        for key in _NESTED_CONFIG_KEYS:
+            sub = cfg.get(key)
+            if isinstance(sub, dict) and isinstance(sub.get("model_type"), str):
+                model_type = sub["model_type"]
+                break
+    if not isinstance(model_type, str):
+        return None
+    for tier in sorted(_TIER_RANK, key = _TIER_RANK.get):
+        if model_type in _config_model_types(tier):
+            return tier
+    return None
+
+
 # --- AutoConfig probe: general tier resolution for ambiguous models ----------
 # When the cheap signals only say "needs some 5.x", parse config.json with the built-in
 # parser in each candidate sidecar (lowest first) instead of guessing. Generalizes beyond
@@ -1210,6 +1324,14 @@ def get_transformers_tier(
                             match,
                         )
                         return tier
+            static = _tier_from_config_mapping(cfg)
+            if static is not None and static != "default":
+                logger.info(
+                    "Transformers tier %s selected for %s (config mapping: model_type absent below)",
+                    static,
+                    model_name,
+                )
+                return static
             local_tc = Path(model_name) / "tokenizer_config.json"
             if _safe_is_file(local_tc) and _check_tokenizer_config_needs_v5(model_name, hf_token):
                 if not probe:
@@ -1269,6 +1391,18 @@ def get_transformers_tier(
             return override
         logger.info("Transformers tier 530 selected for %s (config.json check)", model_name)
         return "530"
+    # _load_config_json (not the cache-only reader) so a config served from the hub
+    # cache during a transient outage still feeds the mapping resolver.
+    remote_cfg = _load_config_json(model_name, hf_token)
+    if remote_cfg is not None:
+        static = _tier_from_config_mapping(remote_cfg)
+        if static is not None and static != "default":
+            logger.info(
+                "Transformers tier %s selected for %s (config mapping: model_type absent below)",
+                static,
+                model_name,
+            )
+            return static
     if _check_tokenizer_config_needs_v5(model_name, hf_token):
         if not probe:
             return "530"
diff --git a/studio/backend/utils/wheel_utils.py b/studio/backend/utils/wheel_utils.py
index 98697df83c..1b5926fd49 100644
--- a/studio/backend/utils/wheel_utils.py
+++ b/studio/backend/utils/wheel_utils.py
@@ -26,11 +26,14 @@ FLASH_ATTN_RELEASE_BASE_URL = "https://github.com/Dao-AILab/flash-attention/rele
 def has_blackwell_gpu() -> bool:
     """Return True if any visible NVIDIA GPU has compute capability >= 10.0 (Blackwell).
 
-    Dao-AILab ships no flash-attention wheels for these archs and older-arch wheels
-    fail to load, so callers use this to skip the flash-attn install path. Cached
-    for the process lifetime; tests mocking nvidia-smi must call
+    Cached for the process lifetime; tests mocking nvidia-smi must call
     ``has_blackwell_gpu.cache_clear()`` first.
     """
+    # Detection disabled for now: Dao-AILab ships Blackwell (sm_100+) flash-attn
+    # wheels and url_exists() already gates resolution, so we no longer skip
+    # flash-attn on Blackwell. The nvidia-smi probe below is kept for possible
+    # future arch-based gating; drop this early return to re-enable it.
+    return False
     exe = shutil.which("nvidia-smi")
     if not exe:
         return False
@@ -117,6 +120,19 @@ def probe_torch_wheel_env(*, timeout: int | None = None) -> dict[str, str] | Non
     return env
 
 
+# torch 2.11 has no native prebuilt wheels for flash-attn / causal-conv1d / mamba
+# yet, but their torch 2.10 CUDA wheels load and pass the projects' own test suites
+# on torch 2.11 (verified on B200: FA2 fwd/bwd, causal-conv1d, and mamba selective
+# scan all match reference). Reuse the torch 2.10 wheels on torch 2.11 so a 2.11
+# install still gets these prebuilt accelerators instead of building from source.
+_PREBUILT_WHEEL_TORCH_MM = {"2.11": "2.10"}
+
+
+def prebuilt_wheel_torch_mm(torch_mm: str) -> str:
+    """Map a torch major.minor to the one whose prebuilt accelerator wheels to use."""
+    return _PREBUILT_WHEEL_TORCH_MM.get(torch_mm, torch_mm)
+
+
 def direct_wheel_url(
     *,
     filename_prefix: str,
@@ -130,7 +146,7 @@ def direct_wheel_url(
 
     filename = (
         f"{filename_prefix}-{package_version}"
-        f"+cu{env['cuda_major']}torch{env['torch_mm']}"
+        f"+cu{env['cuda_major']}torch{prebuilt_wheel_torch_mm(env['torch_mm'])}"
         f"cxx11abi{env['cxx11abi']}-{env['python_tag']}-{env['python_tag']}"
         f"-{env['platform_tag']}.whl"
     )
@@ -152,7 +168,7 @@ def flash_attn_package_version(torch_mm: str) -> str | None:
 def flash_attn_wheel_url(env: dict[str, str] | None) -> str | None:
     if env is None:
         return None
-    package_version = flash_attn_package_version(env["torch_mm"])
+    package_version = flash_attn_package_version(prebuilt_wheel_torch_mm(env["torch_mm"]))
     if package_version is None:
         return None
     return direct_wheel_url(
diff --git a/studio/frontend/src/components/app-sidebar.tsx b/studio/frontend/src/components/app-sidebar.tsx
index f59a952b3a..1a74b38524 100644
--- a/studio/frontend/src/components/app-sidebar.tsx
+++ b/studio/frontend/src/components/app-sidebar.tsx
@@ -81,7 +81,6 @@ import {
   TestTube01Icon,
   ZapIcon,
 } from "@hugeicons/core-free-icons";
-import { listStoredChatThreads } from "@/features/chat/utils/chat-history-storage";
 import {
   Tooltip,
   TooltipContent,
@@ -97,6 +96,7 @@ import {
   createChatProject,
   deleteChatProject,
   deleteChatItem,
+  listStoredChatThreads,
   moveChatItemToProject,
   renameChatItem,
   renameChatProject,
@@ -582,7 +582,14 @@ export function AppSidebar() {
   useEffect(() => {
     if (!pendingRename) return;
     const match = allChatItems.find((i) => i.id === pendingRename.id);
-    if (match && match.title === pendingRename.title) setPendingRename(null);
+    if (!match || match.title !== pendingRename.title) return;
+    queueMicrotask(() => {
+      setPendingRename((current) =>
+        current?.id === pendingRename.id && current.title === pendingRename.title
+          ? null
+          : current,
+      );
+    });
   }, [allChatItems, pendingRename]);
   const [creatingProject, setCreatingProject] = useState(false);
   const [projectNameDraft, setProjectNameDraft] = useState("");
@@ -680,12 +687,6 @@ export function AppSidebar() {
     useState<DeleteTarget | null>(null);
   const [deleteProjectFiles, setDeleteProjectFiles] = useState(false);
 
-  useEffect(() => {
-    if (confirmingDelete?.kind !== "project") {
-      setDeleteProjectFiles(false);
-    }
-  }, [confirmingDelete]);
-
   async function commitDelete() {
     const target = confirmingDelete;
     if (!target) return;
@@ -1572,9 +1573,6 @@ export function AppSidebar() {
                   >
                     <HugeiconsIcon icon={Globe02Icon} strokeWidth={1.75} className="size-[18px]" />
                     <span>{t("shell.navigation.api")}</span>
-                    <span className="ml-auto rounded-full bg-emerald-500/10 px-2 py-1 text-[10px] leading-none font-semibold text-emerald-700 dark:text-emerald-300">
-                      {t("common.new")}
-                    </span>
                   </DropdownMenuItem>
                   <DropdownMenuItem
                     ref={anchorRef as React.Ref<HTMLDivElement>}
diff --git a/studio/frontend/src/components/assistant-ui/message-response-details-sheet.tsx b/studio/frontend/src/components/assistant-ui/message-response-details-sheet.tsx
index 823696693a..331a06a4c6 100644
--- a/studio/frontend/src/components/assistant-ui/message-response-details-sheet.tsx
+++ b/studio/frontend/src/components/assistant-ui/message-response-details-sheet.tsx
@@ -18,7 +18,7 @@ import {
   useExternalProvidersStore,
 } from "@/features/chat";
 import { cn } from "@/lib/utils";
-import { FileDatabaseIcon } from "@hugeicons/core-free-icons";
+import { HelpCircleIcon } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
 import { useMessage, useMessageTiming } from "@assistant-ui/react";
 import type { FC, ReactNode } from "react";
@@ -341,7 +341,7 @@ export const MessageResponseDetailsSheet: FC<{
         <SheetHeader className="border-b p-4">
           <SheetTitle className="flex items-center gap-2 pr-10 font-heading text-base">
             <HugeiconsIcon
-              icon={FileDatabaseIcon}
+              icon={HelpCircleIcon}
               strokeWidth={1.75}
               className="size-icon text-chat-icon-fg"
             />
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx b/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
index 9890c5f574..ff0351883d 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
+++ b/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
@@ -393,6 +393,7 @@ function ModelRow({
   vramEst,
   gpuGb,
   tooltipText,
+  hubUrl,
   optionProps,
   onArrowDownIntoChildren,
   capabilities,
@@ -409,6 +410,10 @@ function ModelRow({
   vramEst?: number;
   gpuGb?: number;
   tooltipText?: ReactNode;
+  /** Hugging Face address (e.g. "huggingface.co/owner/name") for online/Hub
+   * rows; surfaced on hover so their repo id / URL is discoverable the same
+   * way local rows show an on-disk path. Omit to show no address line. */
+  hubUrl?: string;
   optionProps?: ModelRowOptionProps;
   onArrowDownIntoChildren?: () => boolean;
   /** Capability override (HF rows have tags); falls back to name detection. */
@@ -546,30 +551,41 @@ function ModelRow({
     </button>
   );
 
-  if (vramTooltipText) {
-    return (
-      <Tooltip>
-        <TooltipTrigger asChild={true}>{content}</TooltipTrigger>
-        <TooltipContent
-          side="left"
-          className="tooltip-compact max-w-xs break-all"
-        >
-          {label}
-          <span className="block text-[10px] mt-1">{vramTooltipText}</span>
-        </TooltipContent>
-      </Tooltip>
-    );
-  }
+  // Optional Hugging Face address line for online/Hub rows, rendered under
+  // whichever tooltip shows so the repo id / URL is always visible on hover.
+  const hubUrlLine = hubUrl ? (
+    <span className="block mt-1 text-[10px] text-muted-foreground break-all">
+      {hubUrl}
+    </span>
+  ) : null;
 
-  if (tooltipText) {
+  const tooltipBody = vramTooltipText ? (
+    <>
+      {label}
+      <span className="block text-[10px] mt-1">{vramTooltipText}</span>
+      {hubUrlLine}
+    </>
+  ) : tooltipText ? (
+    <>
+      {tooltipText}
+      {hubUrlLine}
+    </>
+  ) : hubUrl ? (
+    <>
+      <span className="block break-words">{label}</span>
+      {hubUrlLine}
+    </>
+  ) : null;
+
+  if (tooltipBody) {
     return (
-      <Tooltip>
+      <Tooltip delayDuration={700}>
         <TooltipTrigger asChild={true}>{content}</TooltipTrigger>
         <TooltipContent
           side="left"
           className="tooltip-compact max-w-xs break-all"
         >
-          {tooltipText}
+          {tooltipBody}
         </TooltipContent>
       </Tooltip>
     );
@@ -1193,6 +1209,13 @@ function localPathTooltip(name: string, path: string): ReactNode {
   );
 }
 
+/** Hugging Face address for an online/Hub row, or undefined when the repo id is
+ * missing so the row shows no (empty) address line on hover. */
+function hubRepoUrl(id: string | null | undefined): string | undefined {
+  const trimmed = id?.trim();
+  return trimmed ? `huggingface.co/${trimmed}` : undefined;
+}
+
 /** Whether a local model is an MLX build (name hint). MLX runs on Mac only, so
  * callers gate visibility on the host being a Mac. */
 function localModelIsMlx(m: LocalModelInfo): boolean {
@@ -2462,6 +2485,7 @@ export function HubModelPicker({
           <div className="min-w-0 flex-1">
             <ModelRow
               label={c.repo_id}
+              tooltipText={localPathTooltip(c.repo_id, c.cache_path)}
               meta="GGUF"
               showVision={c.has_vision ?? visionByRepo[c.repo_id]}
               selected={isSelected}
@@ -2518,6 +2542,7 @@ export function HubModelPicker({
         <div className="min-w-0 flex-1">
           <ModelRow
             label={c.repo_id}
+            hubUrl={hubRepoUrl(c.repo_id)}
             meta={`${isMlxId(c.repo_id) ? "MLX" : "Safetensors"} · ${formatBytes(
               c.size_bytes,
             )}`}
@@ -3378,6 +3403,7 @@ export function HubModelPicker({
                         <div key={id}>
                           <ModelRow
                             label={id}
+                            hubUrl={hubRepoUrl(id)}
                             hideOwner={true}
                             downloaded={downloadedSet.has(id.toLowerCase())}
                             capabilities={capsById.get(id)}
@@ -3463,6 +3489,7 @@ export function HubModelPicker({
                       <div key={id}>
                         <ModelRow
                           label={id}
+                          hubUrl={hubRepoUrl(id)}
                           capabilities={capsById.get(id)}
                           meta={
                             isKnownGgufRepo(id)
@@ -3545,6 +3572,7 @@ export function HubModelPicker({
                         <div key={id}>
                           <ModelRow
                             label={id}
+                            hubUrl={hubRepoUrl(id)}
                             capabilities={capsById.get(id)}
                             meta={
                               isSearchGguf
diff --git a/studio/frontend/src/components/assistant-ui/thread.tsx b/studio/frontend/src/components/assistant-ui/thread.tsx
index 09551cd413..d987092c48 100644
--- a/studio/frontend/src/components/assistant-ui/thread.tsx
+++ b/studio/frontend/src/components/assistant-ui/thread.tsx
@@ -127,6 +127,7 @@ import {
   FileDatabaseIcon,
   Folder01Icon,
   FolderAddIcon,
+  HelpCircleIcon,
   Image03Icon,
   McpServerIcon,
   PencilRulerIcon,
@@ -3952,7 +3953,7 @@ const AssistantActionBar: FC = () => {
                   strokeWidth={1.75}
                   className="size-icon"
                 />
-                Export as Markdown
+                Export as markdown
               </ActionBarMorePrimitive.Item>
             </ActionBarPrimitive.ExportMarkdown>
             <ActionBarMorePrimitive.Item
@@ -3960,7 +3961,7 @@ const AssistantActionBar: FC = () => {
               className="aui-action-bar-more-item flex cursor-pointer select-none items-center gap-2 rounded-[12px] px-3 py-2 text-sm outline-none hover:bg-accent hover:text-accent-foreground focus:bg-accent focus:text-accent-foreground"
             >
               <HugeiconsIcon
-                icon={FileDatabaseIcon}
+                icon={HelpCircleIcon}
                 strokeWidth={1.75}
                 className="size-icon"
               />
diff --git a/studio/frontend/src/components/floating-monitor.tsx b/studio/frontend/src/components/floating-monitor.tsx
index bce4bf2831..0a51875de9 100644
--- a/studio/frontend/src/components/floating-monitor.tsx
+++ b/studio/frontend/src/components/floating-monitor.tsx
@@ -3,27 +3,35 @@
 
 import { Button } from "@/components/ui/button";
 import { Progress } from "@/components/ui/progress";
-import { useMonitorOverlayStore } from "@/features/settings/stores/monitor-overlay-store";
+import { useMonitorOverlayStore } from "@/features/settings";
 import { useSystemInfo } from "@/hooks/use-system";
 import { useT } from "@/i18n";
 import { cn } from "@/lib/utils";
 import { CpuIcon, GripVerticalIcon, XIcon } from "lucide-react";
-import { motion } from "motion/react";
-import { useRef } from "react";
+import { AnimatePresence, motion, useDragControls } from "motion/react";
+import { type PointerEvent, useMemo, useState } from "react";
 
 function clampPercent(value: number): number {
   return Math.max(0, Math.min(100, value));
 }
 
 function usageIndicatorClass(percent: number): string {
-  if (percent >= 90) return "bg-destructive";
-  if (percent >= 70) return "bg-amber-500";
+  if (percent >= 90) {
+    return "bg-destructive";
+  }
+  if (percent >= 70) {
+    return "bg-amber-500";
+  }
   return "bg-primary";
 }
 
 function usageTextClass(percent: number): string {
-  if (percent >= 90) return "text-destructive";
-  if (percent >= 70) return "text-amber-600 dark:text-amber-400";
+  if (percent >= 90) {
+    return "text-destructive";
+  }
+  if (percent >= 70) {
+    return "text-amber-600 dark:text-amber-400";
+  }
   return "text-primary";
 }
 
@@ -39,9 +47,18 @@ export function FloatingMonitor() {
   const { isOpen, setIsOpen } = useMonitorOverlayStore();
   const systemInfo = useSystemInfo({ enabled: isOpen, pollMs: 5000 });
 
-  const constraintsRef = useRef<HTMLDivElement>(null);
+  const [constraintsElement, setConstraintsElement] =
+    useState<HTMLDivElement | null>(null);
+  const constraintsRef = useMemo(
+    () => ({ current: constraintsElement }),
+    [constraintsElement],
+  );
+  const dragControls = useDragControls();
 
-  if (!isOpen) return null;
+  function startDrag(event: PointerEvent<HTMLDivElement>) {
+    event.preventDefault();
+    dragControls.start(event);
+  }
 
   const ramTotal = systemInfo.memory?.total_gb ?? 0;
   const ramAvailable = systemInfo.memory?.available_gb ?? 0;
@@ -64,99 +81,109 @@ export function FloatingMonitor() {
   const hasGpu = (systemInfo.gpu?.available ?? false) && devices.length > 0;
 
   return (
-    <div
-      ref={constraintsRef}
-      className="fixed inset-0 z-50 pointer-events-none"
-    >
-      <motion.div
-        layout={true}
-        drag={true}
-        dragConstraints={constraintsRef}
-        dragElastic={0.1}
-        dragMomentum={false}
-        initial={{ opacity: 0, scale: 0.9 }}
-        animate={{ opacity: 1, scale: 1 }}
-        exit={{ opacity: 0, scale: 0.9 }}
-        className="settings-surface fixed bottom-4 right-4 w-64 max-w-[calc(100vw-2rem)] resize overflow-hidden rounded-xl border border-border/70 p-3 shadow-border ring-0 backdrop-blur-sm pointer-events-auto cursor-default select-none"
-      >
-        <div className="mb-2 flex items-center justify-between gap-2 border-b border-border/60 pb-2">
-          <div className="flex min-w-0 flex-1 items-center gap-1.5 truncate text-xs font-semibold text-foreground">
-            <CpuIcon className="size-3.5 shrink-0 text-primary" />
-            <span className="truncate">
-              {t("settings.resources.liveMonitor.title")}
-            </span>
-          </div>
-          <div className="flex items-center gap-1 shrink-0">
-            <div className="cursor-grab rounded-md px-1 text-muted-foreground/60 transition-colors hover:bg-muted/60 hover:text-muted-foreground active:cursor-grabbing">
-              <GripVerticalIcon className="size-3.5" />
-            </div>
-
-            <Button
-              size="icon-xs"
-              variant="ghost"
-              className="text-muted-foreground hover:text-foreground"
-              onClick={() => setIsOpen(false)}
-              title={t("common.close")}
-              aria-label={t("common.close")}
-            >
-              <XIcon className="size-3" />
-            </Button>
-          </div>
-        </div>
-
-        <motion.div
-          initial={{ opacity: 0, height: 0 }}
-          animate={{ opacity: 1, height: "auto" }}
-          exit={{ opacity: 0, height: 0 }}
-          className="space-y-3 overflow-hidden"
+    <AnimatePresence>
+      {isOpen && (
+        <div
+          ref={setConstraintsElement}
+          className="fixed inset-0 z-50 pointer-events-none"
         >
-          <div className="space-y-1">
-            <div className="flex justify-between text-[11px] font-medium font-mono">
-              <span>{t("settings.resources.liveMonitor.ram")}</span>
-              <span className={cn("tabular-nums", usageTextClass(ramPercent))}>
-                {Math.round(ramPercent)}%
-              </span>
-            </div>
-            <div className="text-xs text-muted-foreground font-mono tabular-nums">
-              {formatGiB(ramUsed)} / {formatGiB(ramTotal)}
-            </div>
-            <Progress
-              value={ramPercent}
-              className="mt-1 h-1.5 rounded-full bg-muted"
-              indicatorClassName={usageIndicatorClass(ramPercent)}
-            />
-          </div>
-
-          {hasGpu && (
-            <div className="space-y-1">
-              <div className="flex justify-between text-[11px] font-medium font-mono">
-                <span className="truncate flex-1 pr-2">
-                  {t("settings.resources.liveMonitor.vram")}{" "}
-                  {devices.length > 1
-                    ? `(${devices.length} GPUs)`
-                    : `(${devices[0].name ?? "GPU"})`}
+          <motion.div
+            drag={true}
+            dragControls={dragControls}
+            dragListener={false}
+            dragConstraints={constraintsRef}
+            dragElastic={0}
+            dragMomentum={false}
+            initial={{ opacity: 0, scale: 0.9 }}
+            animate={{ opacity: 1, scale: 1 }}
+            exit={{ opacity: 0, scale: 0.9 }}
+            className="settings-surface fixed bottom-4 right-4 w-64 max-w-[calc(100vw-2rem)] resize overflow-hidden rounded-xl border border-border/70 p-3 shadow-border ring-0 backdrop-blur-sm pointer-events-auto cursor-default select-none"
+          >
+            <div className="mb-2 flex items-center justify-between gap-2 border-b border-border/60 pb-2">
+              <div className="flex min-w-0 flex-1 items-center gap-1.5 truncate text-xs font-semibold text-foreground">
+                <CpuIcon className="size-3.5 shrink-0 text-primary" />
+                <span className="truncate">
+                  {t("settings.resources.liveMonitor.title")}
                 </span>
-                <span
-                  className={cn(
-                    "shrink-0 tabular-nums",
-                    usageTextClass(vramPercent),
-                  )}
+              </div>
+              <div className="flex items-center gap-1 shrink-0">
+                <div
+                  onPointerDown={startDrag}
+                  className="touch-none cursor-grab rounded-md px-1 text-muted-foreground/60 transition-colors hover:bg-muted/60 hover:text-muted-foreground active:cursor-grabbing"
                 >
-                  {Math.round(vramPercent)}%
-                </span>
+                  <GripVerticalIcon className="size-3.5" />
+                </div>
+
+                <Button
+                  size="icon-xs"
+                  variant="ghost"
+                  className="text-muted-foreground hover:text-foreground"
+                  onClick={() => setIsOpen(false)}
+                  title={t("common.close")}
+                  aria-label={t("common.close")}
+                >
+                  <XIcon className="size-3" />
+                </Button>
               </div>
-              <div className="text-xs text-muted-foreground font-mono tabular-nums">
-                {formatGiB(vramUsed)} / {formatGiB(vramTotal)}
-              </div>
-              <Progress
-                value={vramPercent}
-                className="mt-1 h-1.5 rounded-full bg-muted"
-                indicatorClassName={usageIndicatorClass(vramPercent)}
-              />
             </div>
-          )}
-        </motion.div>
-      </motion.div>
-    </div>
+
+            <motion.div
+              initial={{ opacity: 0, height: 0 }}
+              animate={{ opacity: 1, height: "auto" }}
+              exit={{ opacity: 0, height: 0 }}
+              className="space-y-3 overflow-hidden"
+            >
+              <div className="space-y-1">
+                <div className="flex justify-between text-[11px] font-medium font-mono">
+                  <span>{t("settings.resources.liveMonitor.ram")}</span>
+                  <span
+                    className={cn("tabular-nums", usageTextClass(ramPercent))}
+                  >
+                    {Math.round(ramPercent)}%
+                  </span>
+                </div>
+                <div className="text-xs text-muted-foreground font-mono tabular-nums">
+                  {formatGiB(ramUsed)} / {formatGiB(ramTotal)}
+                </div>
+                <Progress
+                  value={ramPercent}
+                  className="mt-1 h-1.5 rounded-full bg-muted"
+                  indicatorClassName={usageIndicatorClass(ramPercent)}
+                />
+              </div>
+
+              {hasGpu && (
+                <div className="space-y-1">
+                  <div className="flex justify-between text-[11px] font-medium font-mono">
+                    <span className="truncate flex-1 pr-2">
+                      {t("settings.resources.liveMonitor.vram")}{" "}
+                      {devices.length > 1
+                        ? `(${devices.length} GPUs)`
+                        : `(${devices[0].name ?? "GPU"})`}
+                    </span>
+                    <span
+                      className={cn(
+                        "shrink-0 tabular-nums",
+                        usageTextClass(vramPercent),
+                      )}
+                    >
+                      {Math.round(vramPercent)}%
+                    </span>
+                  </div>
+                  <div className="text-xs text-muted-foreground font-mono tabular-nums">
+                    {formatGiB(vramUsed)} / {formatGiB(vramTotal)}
+                  </div>
+                  <Progress
+                    value={vramPercent}
+                    className="mt-1 h-1.5 rounded-full bg-muted"
+                    indicatorClassName={usageIndicatorClass(vramPercent)}
+                  />
+                </div>
+              )}
+            </motion.div>
+          </motion.div>
+        </div>
+      )}
+    </AnimatePresence>
   );
 }
diff --git a/studio/frontend/src/features/chat/api/chat-adapter.ts b/studio/frontend/src/features/chat/api/chat-adapter.ts
index 2ca6ceddf7..ca062bf93c 100644
--- a/studio/frontend/src/features/chat/api/chat-adapter.ts
+++ b/studio/frontend/src/features/chat/api/chat-adapter.ts
@@ -2,6 +2,10 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import { getAuthToken } from "@/features/auth";
+import {
+  loadRememberedLoadSettings,
+  rememberedLoadSettingsKey,
+} from "@/components/assistant-ui/model-selector/remembered-load-settings";
 import { projectHasSources } from "@/features/rag/api/rag-api";
 import { apiUrl } from "@/lib/api-base";
 import { parseParamCountB } from "@/lib/model-size";
@@ -63,11 +67,17 @@ import {
   listStoredChatThreads,
   updateStoredChatThread,
 } from "../utils/chat-history-storage";
+import {
+  readLastLocalModelLoad,
+  recordLastLocalModelLoad,
+  type LastLocalModelKind,
+} from "../utils/last-local-model-load";
 import { getImageInputUnavailableReason } from "../utils/image-input-support";
 import {
   hasClosedThinkTag,
   parseAssistantContent,
 } from "../utils/parse-assistant-content";
+import { resolveLoadMaxSeqLength } from "../presets/preset-policy";
 import {
   generateAudio,
   listCachedGguf,
@@ -1309,6 +1319,30 @@ const BIG_ENDIAN_GGUF_FILENAME_RE = /(^|[-_])be(?:[._-]|$)/gi;
 const GGUF_KNOWN_QUANT_RE =
   /(UD-)?(MXFP[0-9]+(?:_[A-Z0-9]+)*|IQ[0-9]+_[A-Z]+(?:_[A-Z0-9]+)?|TQ[0-9]+_[0-9]+|Q[0-9]+_K_[A-Z]+|Q[0-9]+_[0-9]+|Q[0-9]+_K|BF16|F16|F32)/i;
 
+type AutoLoadCandidate = {
+  id: string;
+  kind: LastLocalModelKind;
+  ggufVariant: string | null;
+  maxSeqLength: number;
+  successLabel: string;
+};
+
+function autoLoadCandidateKey(
+  kind: LastLocalModelKind,
+  id: string,
+  ggufVariant?: string | null,
+): string {
+  return `${kind}:${id.toLowerCase()}:${(ggufVariant ?? "").toLowerCase()}`;
+}
+
+function findCachedRepo<T extends { repo_id: string }>(
+  repos: T[],
+  id: string,
+): T | undefined {
+  const normalized = id.toLowerCase();
+  return repos.find((repo) => repo.repo_id.toLowerCase() === normalized);
+}
+
 function hasBigEndianGgufMarker(filename: string, quant?: string | null): boolean {
   const normalized = filename.replace(/\\/g, "/").toLowerCase();
   const separatorIndex = normalized.lastIndexOf("/");
@@ -1357,14 +1391,18 @@ async function autoLoadSmallestModel(): Promise<{
   const hfToken = store.hfToken || null;
   const trustRemoteCode = store.params.trustRemoteCode ?? false;
   const specSettings = resolveSpeculativeSettingsForLoad();
+  const lastLoaded = readLastLocalModelLoad();
   const toastId = toast("Loading a model…", {
-    description: "Auto-selecting the smallest downloaded model.",
+    description: lastLoaded
+      ? "Loading last used model."
+      : "Auto-selecting the smallest downloaded model.",
     duration: 5000,
     closeButton: true,
   });
   let blockedByTrustRemoteCode = false;
   let hadNonTrustFailure = false;
   let loadAttempts = 0;
+  const skippedAutoLoadCandidates = new Set<string>();
 
   async function canAutoLoad(payload: {
     model_path: string;
@@ -1389,12 +1427,224 @@ async function autoLoadSmallestModel(): Promise<{
     }
     return true;
   }
+
+  async function loadAutoLoadCandidate(
+    candidate: AutoLoadCandidate,
+  ): Promise<boolean> {
+    if (loadAttempts >= MAX_AUTO_LOAD_ATTEMPTS) {
+      return false;
+    }
+    const currentStore = useChatRuntimeStore.getState();
+    const remembered = loadRememberedLoadSettings(
+      rememberedLoadSettingsKey({
+        id: candidate.id,
+        ggufVariant: candidate.ggufVariant,
+      }),
+    );
+    const effectiveMaxSeqLength = resolveLoadMaxSeqLength({
+      modelId: candidate.id,
+      ggufVariant: candidate.ggufVariant,
+      isGguf: candidate.kind === "gguf",
+      customContextLength: remembered?.contextLength ?? null,
+      ggufContextLength: null,
+      currentCheckpoint: currentStore.params.checkpoint,
+      activeGgufVariant: currentStore.activeGgufVariant,
+      maxSeqLength: candidate.maxSeqLength,
+      presetSource: currentStore.activePresetSource,
+    });
+    const effectiveSpeculativeType =
+      remembered?.speculativeType ?? specSettings.speculativeType;
+    const effectiveSpecDraftNMax =
+      remembered?.specDraftNMax ?? specSettings.specDraftNMax;
+    if (
+      !(await canAutoLoad({
+        model_path: candidate.id,
+        max_seq_length: effectiveMaxSeqLength,
+        is_lora: false,
+        gguf_variant: candidate.ggufVariant,
+      }))
+    ) {
+      skippedAutoLoadCandidates.add(
+        autoLoadCandidateKey(candidate.kind, candidate.id, candidate.ggufVariant),
+      );
+      return false;
+    }
+    loadAttempts += 1;
+    const loadResp = await loadModel({
+      model_path: candidate.id,
+      hf_token: hfToken,
+      max_seq_length: effectiveMaxSeqLength,
+      load_in_4bit: true,
+      is_lora: false,
+      gguf_variant: candidate.ggufVariant,
+      trust_remote_code: trustRemoteCode,
+      cache_type_kv: remembered?.kvCacheDtype ?? null,
+      speculative_type: effectiveSpeculativeType,
+      spec_draft_n_max: effectiveSpecDraftNMax,
+      tensor_parallel: remembered?.tensorParallel ?? false,
+    });
+    saveSpeculativeType(effectiveSpeculativeType);
+    useChatRuntimeStore
+      .getState()
+      .setCheckpoint(candidate.id, candidate.ggufVariant ?? undefined);
+    const store = useChatRuntimeStore.getState();
+    store.setModelRequiresTrustRemoteCode(
+      loadResp.requires_trust_remote_code ?? false,
+    );
+    store.setParams({
+      ...store.params,
+      maxTokens:
+        candidate.kind === "gguf"
+          ? loadResp.context_length ?? 131072
+          : effectiveMaxSeqLength,
+    });
+    const autoModel: ChatModelSummary = {
+      id: candidate.id,
+      name: loadResp.display_name ?? candidate.id,
+      isVision: loadResp.is_vision ?? false,
+      isLora: loadResp.is_lora ?? false,
+      isGguf: loadResp.is_gguf ?? candidate.kind === "gguf",
+      isAudio: loadResp.is_audio ?? false,
+      audioType: loadResp.audio_type ?? null,
+      hasAudioInput: loadResp.has_audio_input ?? false,
+    };
+    if (!store.models.some((m) => m.id === candidate.id)) {
+      store.setModels([...store.models, autoModel]);
+    }
+    if (candidate.kind === "gguf") {
+      useChatRuntimeStore.setState({
+        ggufContextLength: loadResp.context_length ?? 131072,
+        ggufMaxContextLength:
+          loadResp.max_context_length ?? loadResp.context_length ?? 131072,
+        ggufNativeContextLength: loadResp.native_context_length ?? null,
+        supportsReasoning: loadResp.supports_reasoning ?? false,
+        reasoningAlwaysOn: loadResp.reasoning_always_on ?? false,
+        reasoningEnabled: loadResp.supports_reasoning ?? false,
+        ...reasoningCapsFromLoad(loadResp),
+        supportsPreserveThinking: loadResp.supports_preserve_thinking ?? false,
+        supportsTools: loadResp.supports_tools ?? false,
+        ...resolveToolsEnabledOnLoad(loadResp.supports_tools ?? false),
+        kvCacheDtype: loadResp.cache_type_kv ?? null,
+        loadedKvCacheDtype: loadResp.cache_type_kv ?? null,
+        tensorParallel: loadResp.tensor_parallel ?? false,
+        loadedTensorParallel: loadResp.tensor_parallel ?? false,
+        defaultChatTemplate: loadResp.chat_template ?? null,
+        chatTemplateOverride: null,
+        loadedChatTemplateOverride: null,
+        loadedIsMultimodal: isMultimodalResponse(loadResp),
+        loadedIsDiffusion: loadResp.is_diffusion ?? false,
+        ...resolveLoadedSpeculativeSettings(loadResp),
+      });
+    } else {
+      useChatRuntimeStore.setState({
+        supportsReasoning: loadResp.supports_reasoning ?? false,
+        reasoningAlwaysOn: loadResp.reasoning_always_on ?? false,
+        reasoningEnabled: loadResp.supports_reasoning ?? false,
+        ...reasoningCapsFromLoad(loadResp),
+        supportsPreserveThinking: loadResp.supports_preserve_thinking ?? false,
+        supportsTools: loadResp.supports_tools ?? false,
+        ...resolveToolsEnabledOnLoad(loadResp.supports_tools ?? false),
+        kvCacheDtype: loadResp.cache_type_kv ?? null,
+        loadedKvCacheDtype: loadResp.cache_type_kv ?? null,
+        tensorParallel: loadResp.tensor_parallel ?? false,
+        loadedTensorParallel: loadResp.tensor_parallel ?? false,
+        defaultChatTemplate: loadResp.chat_template ?? null,
+        chatTemplateOverride: null,
+        loadedChatTemplateOverride: null,
+        ...resolveLoadedSpeculativeSettings(loadResp),
+        loadedIsMultimodal: isMultimodalResponse(loadResp),
+        loadedIsDiffusion: loadResp.is_diffusion ?? false,
+      });
+    }
+    if (!(loadResp.is_lora ?? false)) {
+      recordLastLocalModelLoad({
+        id: candidate.id,
+        kind: candidate.kind,
+        ggufVariant: candidate.ggufVariant,
+      });
+    }
+    toast.success(candidate.successLabel, { id: toastId });
+    return true;
+  }
   try {
     const [ggufRepos, modelRepos] = await Promise.all([
       listCachedGguf().catch(() => []),
       listCachedModels().catch(() => []),
     ]);
 
+    if (lastLoaded) {
+      if (lastLoaded.kind === "gguf") {
+        const repo = findCachedRepo(ggufRepos, lastLoaded.id);
+        if (repo && lastLoaded.ggufVariant) {
+          try {
+            const variants = await listGgufVariants(repo.repo_id);
+            const variant = variants.variants.find(
+              (entry) =>
+                entry.downloaded &&
+                entry.quant?.toLowerCase() ===
+                  lastLoaded.ggufVariant?.toLowerCase() &&
+                isAutoLoadableGgufVariant(entry),
+            );
+            if (variant) {
+              toast("Loading last used model…", {
+                id: toastId,
+                description: `${repo.repo_id} (${variant.quant})`,
+                duration: 5000,
+              });
+              if (
+                await loadAutoLoadCandidate({
+                  id: repo.repo_id,
+                  kind: "gguf",
+                  ggufVariant: variant.quant,
+                  maxSeqLength: 0,
+                  successLabel: `Loaded ${repo.repo_id} (${variant.quant})`,
+                })
+              ) {
+                return { loaded: true, blockedByTrustRemoteCode: false };
+              }
+            }
+          } catch {
+            hadNonTrustFailure = true;
+            skippedAutoLoadCandidates.add(
+              autoLoadCandidateKey("gguf", repo.repo_id, lastLoaded.ggufVariant),
+            );
+          }
+        }
+      } else {
+        const repo = findCachedRepo(modelRepos, lastLoaded.id);
+        if (repo) {
+          try {
+            toast("Loading last used model…", {
+              id: toastId,
+              description: repo.repo_id,
+              duration: 5000,
+            });
+            if (
+              await loadAutoLoadCandidate({
+                id: repo.repo_id,
+                kind: "model",
+                ggufVariant: null,
+                maxSeqLength: store.params.maxSeqLength,
+                successLabel: `Loaded ${repo.repo_id}`,
+              })
+            ) {
+              return { loaded: true, blockedByTrustRemoteCode: false };
+            }
+          } catch {
+            hadNonTrustFailure = true;
+            skippedAutoLoadCandidates.add(
+              autoLoadCandidateKey("model", repo.repo_id),
+            );
+          }
+        }
+      }
+      toast("Loading a model…", {
+        id: toastId,
+        description: "Auto-selecting the smallest downloaded model.",
+        duration: 5000,
+      });
+    }
+
     // GGUF first: smallest-total-size repo, then its smallest variant.
     if (ggufRepos.length > 0) {
       const sorted = [...ggufRepos].sort((a, b) => a.size_bytes - b.size_bytes);
@@ -1408,82 +1658,23 @@ async function autoLoadSmallestModel(): Promise<{
           if (downloaded.length > 0) {
             const variant = downloaded[0];
             if (
-              !(await canAutoLoad({
-                model_path: repo.repo_id,
-                max_seq_length: 0,
-                is_lora: false,
-                gguf_variant: variant.quant,
-              }))
+              skippedAutoLoadCandidates.has(
+                autoLoadCandidateKey("gguf", repo.repo_id, variant.quant),
+              )
             ) {
               continue;
             }
-            loadAttempts += 1;
-            const loadResp = await loadModel({
-              model_path: repo.repo_id,
-              hf_token: hfToken,
-              max_seq_length: 0,
-              load_in_4bit: true,
-              is_lora: false,
-              gguf_variant: variant.quant,
-              trust_remote_code: trustRemoteCode,
-              speculative_type: specSettings.speculativeType,
-              spec_draft_n_max: specSettings.specDraftNMax,
-            });
-            saveSpeculativeType(specSettings.speculativeType);
-            useChatRuntimeStore
-              .getState()
-              .setCheckpoint(repo.repo_id, variant.quant);
-            const store = useChatRuntimeStore.getState();
-            store.setModelRequiresTrustRemoteCode(
-              loadResp.requires_trust_remote_code ?? false,
-            );
-            store.setParams({
-              ...store.params,
-              maxTokens: loadResp.context_length ?? 131072,
-            });
-            // Add to store so the selector shows the name.
-            const autoModel: ChatModelSummary = {
-              id: repo.repo_id,
-              name: loadResp.display_name ?? repo.repo_id,
-              isVision: loadResp.is_vision ?? false,
-              isLora: loadResp.is_lora ?? false,
-              isGguf: loadResp.is_gguf ?? false,
-              isAudio: loadResp.is_audio ?? false,
-              audioType: loadResp.audio_type ?? null,
-              hasAudioInput: loadResp.has_audio_input ?? false,
-            };
-            const existingModels = store.models;
-            if (!existingModels.some((m) => m.id === repo.repo_id)) {
-              store.setModels([...existingModels, autoModel]);
+            if (
+              await loadAutoLoadCandidate({
+                id: repo.repo_id,
+                kind: "gguf",
+                ggufVariant: variant.quant,
+                maxSeqLength: 0,
+                successLabel: `Loaded ${repo.repo_id} (${variant.quant})`,
+              })
+            ) {
+              return { loaded: true, blockedByTrustRemoteCode: false };
             }
-            useChatRuntimeStore.setState({
-              ggufContextLength: loadResp.context_length ?? 131072,
-              ggufMaxContextLength:
-                loadResp.max_context_length ??
-                loadResp.context_length ??
-                131072,
-              supportsReasoning: loadResp.supports_reasoning ?? false,
-              reasoningAlwaysOn: loadResp.reasoning_always_on ?? false,
-              reasoningEnabled: loadResp.supports_reasoning ?? false,
-              ...reasoningCapsFromLoad(loadResp),
-              supportsPreserveThinking:
-                loadResp.supports_preserve_thinking ?? false,
-              supportsTools: loadResp.supports_tools ?? false,
-              ...resolveToolsEnabledOnLoad(loadResp.supports_tools ?? false),
-              kvCacheDtype: loadResp.cache_type_kv ?? null,
-              loadedKvCacheDtype: loadResp.cache_type_kv ?? null,
-              tensorParallel: loadResp.tensor_parallel ?? false,
-              loadedTensorParallel: loadResp.tensor_parallel ?? false,
-              defaultChatTemplate: loadResp.chat_template ?? null,
-              chatTemplateOverride: null,
-              loadedChatTemplateOverride: null,
-              loadedIsMultimodal: isMultimodalResponse(loadResp),
-              ...resolveLoadedSpeculativeSettings(loadResp),
-            });
-            toast.success(`Loaded ${repo.repo_id} (${variant.quant})`, {
-              id: toastId,
-            });
-            return { loaded: true, blockedByTrustRemoteCode: false };
           }
         } catch {
           hadNonTrustFailure = true;
@@ -1501,64 +1692,23 @@ async function autoLoadSmallestModel(): Promise<{
         if (loadAttempts >= MAX_AUTO_LOAD_ATTEMPTS) break;
         try {
           if (
-            !(await canAutoLoad({
-              model_path: repo.repo_id,
-              max_seq_length: 4096,
-              is_lora: false,
-              gguf_variant: null,
-            }))
+            skippedAutoLoadCandidates.has(
+              autoLoadCandidateKey("model", repo.repo_id),
+            )
           ) {
             continue;
           }
-          loadAttempts += 1;
-          const sfLoadResp = await loadModel({
-            model_path: repo.repo_id,
-            hf_token: hfToken,
-            max_seq_length: 4096,
-            load_in_4bit: true,
-            is_lora: false,
-            gguf_variant: null,
-            trust_remote_code: trustRemoteCode,
-            speculative_type: specSettings.speculativeType,
-            spec_draft_n_max: specSettings.specDraftNMax,
-          });
-          saveSpeculativeType(specSettings.speculativeType);
-          useChatRuntimeStore.getState().setCheckpoint(repo.repo_id);
-          const store = useChatRuntimeStore.getState();
-          store.setModelRequiresTrustRemoteCode(
-            sfLoadResp.requires_trust_remote_code ?? false,
-          );
-          store.setParams({ ...store.params, maxTokens: 4096 });
-          useChatRuntimeStore.setState({
-            supportsReasoning: sfLoadResp.supports_reasoning ?? false,
-            reasoningAlwaysOn: sfLoadResp.reasoning_always_on ?? false,
-            reasoningEnabled: sfLoadResp.supports_reasoning ?? false,
-            ...reasoningCapsFromLoad(sfLoadResp),
-            supportsPreserveThinking:
-              sfLoadResp.supports_preserve_thinking ?? false,
-            supportsTools: sfLoadResp.supports_tools ?? false,
-            // Parity with the GGUF branch above.
-            ...resolveToolsEnabledOnLoad(sfLoadResp.supports_tools ?? false),
-            defaultChatTemplate: sfLoadResp.chat_template ?? null,
-            chatTemplateOverride: null,
-            loadedChatTemplateOverride: null,
-            ...resolveLoadedSpeculativeSettings(sfLoadResp),
-          });
-          const sfModel: ChatModelSummary = {
-            id: repo.repo_id,
-            name: sfLoadResp.display_name ?? repo.repo_id,
-            isVision: sfLoadResp.is_vision ?? false,
-            isLora: sfLoadResp.is_lora ?? false,
-            isGguf: sfLoadResp.is_gguf ?? false,
-          };
-          if (!store.models.some((m) => m.id === repo.repo_id)) {
-            store.setModels([...store.models, sfModel]);
+          if (
+            await loadAutoLoadCandidate({
+              id: repo.repo_id,
+              kind: "model",
+              ggufVariant: null,
+              maxSeqLength: 4096,
+              successLabel: `Loaded ${repo.repo_id}`,
+            })
+          ) {
+            return { loaded: true, blockedByTrustRemoteCode: false };
           }
-          useChatRuntimeStore.setState({
-            loadedIsMultimodal: isMultimodalResponse(sfLoadResp),
-          });
-          toast.success(`Loaded ${repo.repo_id}`, { id: toastId });
-          return { loaded: true, blockedByTrustRemoteCode: false };
         } catch {
           hadNonTrustFailure = true;
           continue;
@@ -1650,6 +1800,11 @@ async function autoLoadSmallestModel(): Promise<{
         loadedIsMultimodal: isMultimodalResponse(loadResp),
         ...resolveLoadedSpeculativeSettings(loadResp),
       });
+      recordLastLocalModelLoad({
+        id: "unsloth/Qwen3.5-4B-MTP-GGUF",
+        kind: "gguf",
+        ggufVariant: "UD-Q4_K_XL",
+      });
       toast.success("Loaded Qwen3.5-4B-MTP (UD-Q4_K_XL)", { id: toastId });
       return { loaded: true, blockedByTrustRemoteCode: false };
     } catch {
diff --git a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
index 798c1658f9..0a4342f208 100644
--- a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
+++ b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
@@ -44,6 +44,7 @@ import {
   mergeBackendRecommendedInference,
   resolveLoadMaxSeqLength,
 } from "../presets/preset-policy";
+import { recordLastLocalModelLoad } from "../utils/last-local-model-load";
 import {
   isMultimodalResponse,
 } from "../types/api";
@@ -818,6 +819,23 @@ export function useChatModelRuntime() {
               }
             }
             await refresh({ signal: abortCtrl.signal });
+            if (
+              !isLora &&
+              !(loadResponse.is_lora ?? false) &&
+              !nativePathToken &&
+              !isLocalModelPath(modelId) &&
+              !isExternalModelId(modelId)
+            ) {
+              if (loadResponse.is_gguf || isGguf || ggufVariant) {
+                recordLastLocalModelLoad({
+                  id: modelId,
+                  kind: "gguf",
+                  ggufVariant: ggufVariant ?? null,
+                });
+              } else {
+                recordLastLocalModelLoad({ id: modelId, kind: "model" });
+              }
+            }
             // A successful load owns the shared (pick-unscoped) settings fields,
             // so any surviving stage is stale: the just-loaded pick itself, or a
             // pick queued for a different model mid-load whose knobs this load
diff --git a/studio/frontend/src/features/chat/index.ts b/studio/frontend/src/features/chat/index.ts
index 7cd9611c71..d070ed15de 100644
--- a/studio/frontend/src/features/chat/index.ts
+++ b/studio/frontend/src/features/chat/index.ts
@@ -36,6 +36,7 @@ export { ChatSearchDialog } from "./components/chat-search-dialog";
 export { setTrainingCompareHandoff } from "./lib/training-compare-handoff";
 export type { ProjectRecord } from "./types";
 export { clearAllChats, countAllChats } from "./utils/clear-all-chats";
+export { listStoredChatThreads } from "./utils/chat-history-storage";
 export { ArtifactCard } from "./artifacts/artifact-card";
 export {
   useChatArtifactsStore,
diff --git a/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts b/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts
index 9386650fee..60788c23bd 100644
--- a/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts
+++ b/studio/frontend/src/features/chat/lib/apply-inference-status-to-store.ts
@@ -3,16 +3,19 @@
 
 import { getInferenceStatus } from "../api/chat-api";
 import { mergeBackendRecommendedInference } from "../presets/preset-policy";
+import { clampReasoningEffortToLevels } from "../provider-capabilities";
 import {
   CHAT_REASONING_ENABLED_KEY,
-  loadOptionalBool,
   type ReasoningEffort,
   type ReasoningStyle,
+  loadOptionalBool,
   resolveToolsEnabledOnLoad,
   useChatRuntimeStore,
 } from "../stores/chat-runtime-store";
-import { isMultimodalResponse, type InferenceStatusResponse } from "../types/api";
-import { clampReasoningEffortToLevels } from "../provider-capabilities";
+import {
+  type InferenceStatusResponse,
+  isMultimodalResponse,
+} from "../types/api";
 import type { ChatModelSummary } from "../types/runtime";
 
 type LocalReasoningEffort = Extract<ReasoningEffort, "low" | "medium" | "high">;
@@ -31,7 +34,10 @@ export function normalizeSpeculativeType(
     return "ngram";
   }
   if (s === "mtp+ngram") return "mtp+ngram";
-  const parts = s.split(",").map((p) => p.trim()).filter(Boolean);
+  const parts = s
+    .split(",")
+    .map((p) => p.trim())
+    .filter(Boolean);
   const hasMtp = parts.some((p) => p === "mtp" || p === "draft-mtp");
   const hasNgram = parts.some(
     (p) => p === "ngram" || p === "ngram-mod" || p === "ngram-simple",
@@ -197,6 +203,12 @@ export function applyActiveModelStatusToStore(
     ggufContextLength: currentGgufContextLength,
     ggufMaxContextLength,
     ggufNativeContextLength,
+    // A non-GGUF status must also drop a stale native-path token: without this the
+    // isGguf OR (activeGgufVariant || activeNativePathToken || ggufContextLength)
+    // stays true after switching from a native GGUF to a transformers model, so a
+    // Codex-only detection would auto-select for a model its preflight rejects. A real
+    // GGUF load reports is_gguf: true, so its token is preserved (the load path owns it).
+    ...(status.is_gguf ? {} : { activeNativePathToken: null }),
     modelRequiresTrustRemoteCode: status.requires_trust_remote_code ?? false,
     defaultChatTemplate: nextDefaultChatTemplate,
     loadedIsMultimodal: isMultimodalResponse(status),
@@ -245,7 +257,7 @@ export function applyActiveModelStatusToStore(
     const mid = checkpointId.toLowerCase();
     if (mid.includes("qwen3.5") || mid.includes("qwen3.6")) {
       const sizeMatch = mid.match(/(\d+\.?\d*)\s*b/);
-      if (sizeMatch && parseFloat(sizeMatch[1]) < 9) {
+      if (sizeMatch && Number.parseFloat(sizeMatch[1]) < 9) {
         reasoningDefault = false;
       }
     }
@@ -281,8 +293,7 @@ export async function tryAdoptServerActiveModel(): Promise<boolean> {
   }
 
   // Re-check after the await: keep a checkpoint the user picked meanwhile.
-  const previousCheckpoint =
-    useChatRuntimeStore.getState().params.checkpoint;
+  const previousCheckpoint = useChatRuntimeStore.getState().params.checkpoint;
   if (previousCheckpoint) {
     return true;
   }
diff --git a/studio/frontend/src/features/chat/utils/last-local-model-load.ts b/studio/frontend/src/features/chat/utils/last-local-model-load.ts
new file mode 100644
index 0000000000..099386fbc7
--- /dev/null
+++ b/studio/frontend/src/features/chat/utils/last-local-model-load.ts
@@ -0,0 +1,86 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+export type LastLocalModelKind = "gguf" | "model";
+
+export type LastLocalModelLoad = {
+  id: string;
+  kind: LastLocalModelKind;
+  ggufVariant: string | null;
+  loadedAt: number;
+};
+
+const STORAGE_KEY = "unsloth.last-local-model-load.v1";
+
+function storage(): Storage | null {
+  try {
+    return typeof localStorage === "undefined" ? null : localStorage;
+  } catch {
+    return null;
+  }
+}
+
+function isLastLocalModelKind(value: unknown): value is LastLocalModelKind {
+  return value === "gguf" || value === "model";
+}
+
+export function readLastLocalModelLoad(): LastLocalModelLoad | null {
+  try {
+    const raw = storage()?.getItem(STORAGE_KEY);
+    if (!raw) {
+      return null;
+    }
+    const parsed = JSON.parse(raw) as Partial<LastLocalModelLoad>;
+    if (
+      typeof parsed.id !== "string" ||
+      !parsed.id.trim() ||
+      !isLastLocalModelKind(parsed.kind) ||
+      typeof parsed.loadedAt !== "number"
+    ) {
+      return null;
+    }
+    if (
+      parsed.kind === "gguf" &&
+      (typeof parsed.ggufVariant !== "string" || !parsed.ggufVariant.trim())
+    ) {
+      return null;
+    }
+    return {
+      id: parsed.id,
+      kind: parsed.kind,
+      ggufVariant:
+        typeof parsed.ggufVariant === "string" ? parsed.ggufVariant : null,
+      loadedAt: parsed.loadedAt,
+    };
+  } catch {
+    return null;
+  }
+}
+
+export function recordLastLocalModelLoad(input: {
+  id: string;
+  kind: LastLocalModelKind;
+  ggufVariant?: string | null;
+}): void {
+  const id = input.id.trim();
+  if (!id) {
+    return;
+  }
+  const ggufVariant = input.ggufVariant?.trim() || null;
+  if (input.kind === "gguf" && !ggufVariant) {
+    return;
+  }
+  try {
+    storage()?.setItem(
+      STORAGE_KEY,
+      JSON.stringify({
+        id,
+        kind: input.kind,
+        ggufVariant: input.kind === "gguf" ? ggufVariant : null,
+        loadedAt: Date.now(),
+      } satisfies LastLocalModelLoad),
+    );
+  } catch {
+    // Ignore disabled storage / quota errors; auto-load falls back to size order.
+  }
+}
diff --git a/studio/frontend/src/features/recipe-studio/api/index.ts b/studio/frontend/src/features/recipe-studio/api/index.ts
index 4fe24b1f6d..f4e8167cb3 100644
--- a/studio/frontend/src/features/recipe-studio/api/index.ts
+++ b/studio/frontend/src/features/recipe-studio/api/index.ts
@@ -494,3 +494,13 @@ export async function removeUnstructuredFile(
     throw new Error("Failed to remove file");
   }
 }
+
+export async function removeUnstructuredBlock(blockId: string): Promise<void> {
+  const res = await authFetch(
+    `${DATA_DESIGNER_API_BASE}/seed/unstructured-block/${encodeURIComponent(blockId)}`,
+    { method: "DELETE" },
+  );
+  if (!res.ok && res.status !== 404) {
+    throw new Error("Failed to remove uploaded files");
+  }
+}
diff --git a/studio/frontend/src/features/recipe-studio/dialogs/seed/seed-dialog.tsx b/studio/frontend/src/features/recipe-studio/dialogs/seed/seed-dialog.tsx
index 0ed7eeed75..53f8566090 100644
--- a/studio/frontend/src/features/recipe-studio/dialogs/seed/seed-dialog.tsx
+++ b/studio/frontend/src/features/recipe-studio/dialogs/seed/seed-dialog.tsx
@@ -53,6 +53,11 @@ import {
   inspectSeedDataset,
   inspectSeedUpload,
 } from "../../api";
+import { useRecipeStudioStore } from "../../stores/recipe-studio";
+import {
+  makeUnstructuredUploadUid,
+  resolveUnstructuredUploadBlockId,
+} from "../../utils/config-factories";
 import { resolveImagePreview } from "../../utils/image-preview";
 import type {
   GithubItemType,
@@ -597,6 +602,41 @@ export function SeedDialog({
   const mode = config.seed_source_type ?? "hf";
   const previewEmpty = getPreviewEmptyStateCopy(mode);
 
+  const queueUploadCleanup = useRecipeStudioStore(
+    (state) => state.queueUploadCleanup,
+  );
+
+  // config.id collides across recipes (ids reset to n1 on import); use a
+  // stable per-block uid instead. Generate one synchronously so the first
+  // rendered drop zone cannot upload under a legacy node id.
+  const uploadUid = config.unstructured_upload_uid?.trim() ?? "";
+  const unstructuredFileCount = config.unstructured_file_ids?.length ?? 0;
+  const generatedUploadUidRef = useRef<string | null>(null);
+  if (
+    mode === "unstructured" &&
+    !uploadUid &&
+    unstructuredFileCount === 0 &&
+    generatedUploadUidRef.current === null
+  ) {
+    generatedUploadUidRef.current = makeUnstructuredUploadUid();
+  }
+  const uploadBlockId = resolveUnstructuredUploadBlockId({
+    configId: config.id,
+    uploadUid,
+    generatedUploadUid: generatedUploadUidRef.current,
+    unstructuredFileCount,
+  });
+
+  useEffect(() => {
+    if (mode !== "unstructured") return;
+    if (uploadUid) return;
+    if (unstructuredFileCount > 0) return;
+    const nextUid =
+      generatedUploadUidRef.current ?? makeUnstructuredUploadUid();
+    generatedUploadUidRef.current = nextUid;
+    onUpdate({ unstructured_upload_uid: nextUid });
+  }, [mode, uploadUid, unstructuredFileCount, onUpdate]);
+
   const prevModeRef = useRef(mode);
   useEffect(() => {
     const prevMode = prevModeRef.current;
@@ -720,6 +760,11 @@ export function SeedDialog({
             subset: config.hf_subset?.trim() || undefined,
             preview_size: 10,
           });
+          // Queue the block's upload directory for deletion after the next
+          // save; only uid-namespaced directories qualify (single owner).
+          if (uploadUid && unstructuredFileCount > 0) {
+            queueUploadCleanup(uploadUid);
+          }
           onUpdate({
             hf_path: response.resolved_path,
             seed_columns: response.columns,
@@ -730,6 +775,7 @@ export function SeedDialog({
             hf_split: response.split ?? "",
             hf_subset: response.subset ?? "",
             local_file_name: "",
+            unstructured_upload_uid: "",
             unstructured_file_ids: [],
             unstructured_file_names: [],
             unstructured_file_sizes: [],
@@ -754,6 +800,11 @@ export function SeedDialog({
             content_base64: payload,
             preview_size: 10,
           });
+          // Queue the block's upload directory for deletion after the next
+          // save; only uid-namespaced directories qualify (single owner).
+          if (uploadUid && unstructuredFileCount > 0) {
+            queueUploadCleanup(uploadUid);
+          }
           onUpdate({
             hf_path: response.resolved_path,
             seed_columns: response.columns,
@@ -765,6 +816,7 @@ export function SeedDialog({
             hf_subset: "",
             hf_split: "",
             local_file_name: localFile.name,
+            unstructured_upload_uid: "",
             unstructured_file_ids: [],
             unstructured_file_names: [],
             unstructured_file_sizes: [],
@@ -789,7 +841,7 @@ export function SeedDialog({
 
           const { chunkSize, chunkOverlap } = resolveChunking(config);
           const response = await inspectSeedUpload({
-            block_id: config.id,
+            block_id: uploadBlockId,
             file_ids: fileIds,
             file_names: fileNames,
             preview_size: 10,
@@ -827,7 +879,18 @@ export function SeedDialog({
         setIsInspecting(false);
       }
     },
-    [config, getCurrentLoadKey, localFile, mode, onUpdate, unstructuredFiles],
+    [
+      config,
+      getCurrentLoadKey,
+      localFile,
+      mode,
+      onUpdate,
+      queueUploadCleanup,
+      unstructuredFiles,
+      unstructuredFileCount,
+      uploadBlockId,
+      uploadUid,
+    ],
   );
 
   useEffect(() => {
@@ -997,7 +1060,7 @@ export function SeedDialog({
 
           {mode === "unstructured" && (
             <UnstructuredDropZone
-              blockId={config.id}
+              blockId={uploadBlockId}
               files={unstructuredFiles}
               onFilesChange={handleUnstructuredFilesChange}
               disabled={isInspecting}
diff --git a/studio/frontend/src/features/recipe-studio/dialogs/seed/unstructured-drop-zone.tsx b/studio/frontend/src/features/recipe-studio/dialogs/seed/unstructured-drop-zone.tsx
index 4e9347e02e..bbdf27110f 100644
--- a/studio/frontend/src/features/recipe-studio/dialogs/seed/unstructured-drop-zone.tsx
+++ b/studio/frontend/src/features/recipe-studio/dialogs/seed/unstructured-drop-zone.tsx
@@ -54,11 +54,17 @@ export function UnstructuredDropZone({
 }: UnstructuredDropZoneProps) {
   const inputRef = useRef<HTMLInputElement>(null);
   const filesRef = useRef(files);
+  const blockIdRef = useRef(blockId);
+  const mountedRef = useRef(true);
   const [isDragOver, setIsDragOver] = useState(false);
 
   useEffect(() => {
     filesRef.current = files;
-  }, [files]);
+    blockIdRef.current = blockId;
+  }, [files, blockId]);
+  useEffect(() => () => {
+    mountedRef.current = false;
+  }, []);
 
   const totalSize = files.reduce((sum, f) => sum + f.size, 0);
 
@@ -134,15 +140,32 @@ export function UnstructuredDropZone({
       if (entry.status === "uploading" && entry.abortController) {
         entry.abortController.abort();
       }
-      if (
+      const needsServerRemove =
         entry.id &&
         entry.status === "ok" &&
-        !deletedIdsRef.current.has(entry.id)
-      ) {
-        deletedIdsRef.current.add(entry.id);
-        void removeUnstructuredFile(blockId, entry.id).catch(() => {});
-      }
+        !deletedIdsRef.current.has(entry.id);
       onFilesChange((prev) => prev.filter((_, i) => i !== index));
+      if (!needsServerRemove) return;
+      deletedIdsRef.current.add(entry.id);
+      removeUnstructuredFile(blockId, entry.id).catch(() => {
+        // Skip if the drop zone unmounted or its block changed: the id no
+        // longer belongs here and restoring would leak it into another block.
+        if (!mountedRef.current || blockIdRef.current !== blockId) return;
+        // Still exists server-side (counts toward quota); restore it at its
+        // original position.
+        deletedIdsRef.current.delete(entry.id);
+        onFilesChange((prev) => {
+          const next = [...prev];
+          next.splice(Math.min(index, next.length), 0, {
+            id: entry.id,
+            name: entry.name,
+            size: entry.size,
+            status: "ok",
+            error: "Remove failed — try again",
+          });
+          return next;
+        });
+      });
     },
     [blockId, onFilesChange],
   );
diff --git a/studio/frontend/src/features/recipe-studio/hooks/use-recipe-persistence.ts b/studio/frontend/src/features/recipe-studio/hooks/use-recipe-persistence.ts
index 0417a6dd22..2d91272469 100644
--- a/studio/frontend/src/features/recipe-studio/hooks/use-recipe-persistence.ts
+++ b/studio/frontend/src/features/recipe-studio/hooks/use-recipe-persistence.ts
@@ -4,11 +4,13 @@
 import { useCallback, useEffect, useMemo, useState } from "react";
 import { toastError, toastSuccess } from "@/shared/toast";
 import { normalizeNonEmptyName } from "@/utils";
+import { removeUnstructuredBlock } from "../api";
 import {
   buildSignature,
   copyTextToClipboard,
   formatSavedLabel,
 } from "../executions/execution-helpers";
+import { useRecipeStudioStore } from "../stores/recipe-studio";
 import { importRecipePayload, type RecipeSnapshot } from "../utils/import";
 import type { RecipePayloadResult } from "../utils/payload/types";
 
@@ -72,7 +74,10 @@ function stripApiKeys(value: unknown): unknown {
     !Array.isArray(output.env)
   ) {
     output.env = Object.fromEntries(
-      Object.keys(output.env as Record<string, unknown>).map((envKey) => [envKey, ""]),
+      Object.keys(output.env as Record<string, unknown>).map((envKey) => [
+        envKey,
+        "",
+      ]),
     );
   }
   return output;
@@ -82,10 +87,7 @@ function inferHfRepoIdFromPath(pathValue: unknown): string {
   if (typeof pathValue !== "string") {
     return "";
   }
-  const parts = pathValue
-    .trim()
-    .split("/")
-    .filter(Boolean);
+  const parts = pathValue.trim().split("/").filter(Boolean);
   if (parts.length >= 3 && parts[0] === "datasets") {
     return `${parts[1]}/${parts[2]}`;
   }
@@ -126,8 +128,7 @@ function sanitizeSeedForShare(payload: unknown): unknown {
     typeof ui?.seed_source_type === "string" ? ui.seed_source_type : null;
   const sourceType =
     typeof source?.seed_type === "string" ? source.seed_type : null;
-  const shouldResetHfState =
-    sourceType === "hf" || uiSourceType === "hf";
+  const shouldResetHfState = sourceType === "hf" || uiSourceType === "hf";
   const shouldResetLocalState =
     sourceType === "local" ||
     sourceType === "unstructured" ||
@@ -144,6 +145,7 @@ function sanitizeSeedForShare(payload: unknown): unknown {
       ui.seed_drop_columns = [];
       ui.seed_preview_rows = [];
       ui.local_file_name = "";
+      ui.unstructured_upload_uid = "";
       ui.unstructured_file_ids = [];
       ui.unstructured_file_names = [];
       ui.unstructured_file_sizes = [];
@@ -165,6 +167,7 @@ function sanitizeSeedForShare(payload: unknown): unknown {
       ui.seed_drop_columns = [];
       ui.seed_preview_rows = [];
       ui.local_file_name = "";
+      ui.unstructured_upload_uid = "";
       ui.unstructured_file_ids = [];
       ui.unstructured_file_names = [];
       ui.unstructured_file_sizes = [];
@@ -174,6 +177,43 @@ function sanitizeSeedForShare(payload: unknown): unknown {
   return root;
 }
 
+// Delete queued upload directories once a save stops referencing them, so a
+// reload before autosave can never leave the saved recipe pointing at
+// already-deleted files. Skips any uid the just-saved payload still uses.
+function drainQueuedUploadCleanups(
+  savedPayload: RecipePayloadResult["payload"],
+): void {
+  const pending = useRecipeStudioStore.getState().pendingUploadCleanups;
+  if (pending.length === 0) {
+    return;
+  }
+  const ui =
+    savedPayload && typeof savedPayload === "object"
+      ? (savedPayload as { ui?: Record<string, unknown> }).ui
+      : undefined;
+  const savedUid =
+    ui && typeof ui.unstructured_upload_uid === "string"
+      ? ui.unstructured_upload_uid
+      : "";
+  const ready = pending.filter((uid) => uid !== savedUid);
+  if (ready.length === 0) {
+    return;
+  }
+  for (const uid of ready) {
+    void removeUnstructuredBlock(uid)
+      .then(() => {
+        useRecipeStudioStore.setState((state) => ({
+          pendingUploadCleanups: state.pendingUploadCleanups.filter(
+            (pendingUid) => pendingUid !== uid,
+          ),
+        }));
+      })
+      .catch((error) => {
+        console.warn("Failed to clean up uploaded documents:", error);
+      });
+  }
+}
+
 export function useRecipePersistence({
   recipeId,
   initialRecipeName,
@@ -202,8 +242,10 @@ export function useRecipePersistence({
     () => buildSignature(normalizedWorkflowName, currentPayload),
     [currentPayload, normalizedWorkflowName],
   );
-  const isDirty = savedSignature.length > 0 && currentSignature !== savedSignature;
-  const saveTone: SaveTone = !isDirty && Boolean(lastSavedAt) ? "success" : "error";
+  const isDirty =
+    savedSignature.length > 0 && currentSignature !== savedSignature;
+  const saveTone: SaveTone =
+    !isDirty && Boolean(lastSavedAt) ? "success" : "error";
   const savedAtLabel = formatSavedLabel(lastSavedAt);
 
   useEffect(() => {
@@ -214,7 +256,9 @@ export function useRecipePersistence({
     setLastSavedAt(initialSavedAt);
     setCopied(false);
 
-    const parsed = importRecipePayload(JSON.stringify(initialPayload));
+    const parsed = importRecipePayload(JSON.stringify(initialPayload), {
+      preserveUnstructuredUploads: true,
+    });
     if (parsed.snapshot) {
       loadRecipe(parsed.snapshot);
     } else {
@@ -252,6 +296,7 @@ export function useRecipePersistence({
       });
       setLastSavedAt(result.updatedAt);
       setSavedSignature(buildSignature(nextName, currentPayload));
+      drainQueuedUploadCleanups(currentPayload);
     } catch (error) {
       console.error("Save recipe failed:", error);
       toastError("Save failed", "Could not save recipe.");
@@ -270,11 +315,28 @@ export function useRecipePersistence({
     return () => window.clearTimeout(timeoutId);
   }, [isDirty, persistRecipe, saveLoading]);
 
+  // Drain queued cleanups even when autosave is skipped: a net-zero edit (add
+  // then remove an unstructured seed before the 800ms debounce) keeps isDirty
+  // false, so the autosave effect never drains and the queued uid leaks its
+  // upload dir. Not-dirty means currentPayload equals the saved recipe, and
+  // drain skips the uid it still references, so only dirs no saved recipe
+  // points at are deleted (keeps the save-first invariant).
+  useEffect(() => {
+    if (!initialRecipeReady || isDirty || saveLoading) {
+      return;
+    }
+    drainQueuedUploadCleanups(currentPayload);
+  }, [currentPayload, initialRecipeReady, isDirty, saveLoading]);
+
   const copyRecipe = useCallback(async (): Promise<void> => {
     setCopied(false);
     try {
-      const safePayload = sanitizeSeedForShare(stripApiKeys(payloadResult.payload));
-      const ok = await copyTextToClipboard(JSON.stringify(safePayload, null, 2));
+      const safePayload = sanitizeSeedForShare(
+        stripApiKeys(payloadResult.payload),
+      );
+      const ok = await copyTextToClipboard(
+        JSON.stringify(safePayload, null, 2),
+      );
       if (!ok) {
         throw new Error("Clipboard not available.");
       }
diff --git a/studio/frontend/src/features/recipe-studio/stores/recipe-studio.ts b/studio/frontend/src/features/recipe-studio/stores/recipe-studio.ts
index 8659cad4fb..a1ff72ee9b 100644
--- a/studio/frontend/src/features/recipe-studio/stores/recipe-studio.ts
+++ b/studio/frontend/src/features/recipe-studio/stores/recipe-studio.ts
@@ -36,6 +36,7 @@ import {
 } from "../utils/handles";
 import type { RecipeSnapshot } from "../utils/import";
 import { getLayoutedElements } from "../utils/layout";
+import { makeUnstructuredUploadUid } from "../utils/config-factories";
 import {
   centerModelInfraNodes,
   optimizeModelInfraEdgeHandles,
@@ -76,6 +77,12 @@ type RecipeStudioState = {
   nextId: number;
   nextY: number;
   fitViewTick: number;
+  // Upload-uid directories whose owning block dropped them; server-side
+  // deletion is deferred until a save no longer references them, so a
+  // reload before autosave cannot leave a saved recipe pointing at
+  // deleted files.
+  pendingUploadCleanups: string[];
+  queueUploadCleanup: (uid: string) => void;
   setSheetOpen: (open: boolean) => void;
   setSheetView: (view: SheetView) => void;
   setProcessors: (processors: RecipeProcessorConfig[]) => void;
@@ -137,6 +144,7 @@ const INITIAL_STATE = {
   nextId: 3,
   nextY: 280,
   fitViewTick: 0,
+  pendingUploadCleanups: [],
 } satisfies Pick<
   RecipeStudioState,
   | "nodes"
@@ -154,6 +162,7 @@ const INITIAL_STATE = {
   | "nextId"
   | "nextY"
   | "fitViewTick"
+  | "pendingUploadCleanups"
 >;
 
 function buildAddedNodeState(
@@ -269,6 +278,20 @@ function isModelSemanticEdge(
   );
 }
 
+// Upload uid of a seed block whose server-side directory becomes orphaned
+// when the block drops it. Only uid directories qualify (single owner);
+// legacy node-id directories can be shared by other recipes.
+function seedUploadCleanupUid(config: NodeConfig | undefined): string | null {
+  if (!config || config.kind !== "seed") {
+    return null;
+  }
+  const uid = config.unstructured_upload_uid?.trim();
+  if (!uid || !config.unstructured_file_ids?.length) {
+    return null;
+  }
+  return uid;
+}
+
 export const useRecipeStudioStore = create<RecipeStudioState>((set, get) => ({
   ...INITIAL_STATE,
   setSheetOpen: (open) => set({ sheetOpen: open }),
@@ -278,6 +301,12 @@ export const useRecipeStudioStore = create<RecipeStudioState>((set, get) => ({
   setDialogOpen: (open) => set({ dialogOpen: open }),
   setExecutionLocked: (locked) => set({ executionLocked: locked }),
   resetRecipe: () => set(INITIAL_STATE),
+  queueUploadCleanup: (uid) =>
+    set((state) =>
+      state.pendingUploadCleanups.includes(uid)
+        ? state
+        : { pendingUploadCleanups: [...state.pendingUploadCleanups, uid] },
+    ),
   selectConfig: (id) => set({ activeConfigId: id, dialogOpen: false }),
   openConfig: (id) => set({ activeConfigId: id, dialogOpen: true }),
   setLayoutDirection: (direction) =>
@@ -383,7 +412,18 @@ export const useRecipeStudioStore = create<RecipeStudioState>((set, get) => ({
       }
       return buildAddedNodeState(state, "sampler", type, position, openDialog);
     }),
-  addSeedNode: (type, position, openDialog = true) =>
+  addSeedNode: (type, position, openDialog = true) => {
+    const current = get();
+    if (!current.executionLocked) {
+      // The reset below clears the block's upload uid and file list; queue
+      // its server-side directory for deletion after the next save.
+      const uid = seedUploadCleanupUid(
+        Object.values(current.configs).find((config) => config.kind === "seed"),
+      );
+      if (uid) {
+        current.queueUploadCleanup(uid);
+      }
+    }
     set((state) => {
       if (state.executionLocked) {
         return state;
@@ -413,6 +453,8 @@ export const useRecipeStudioStore = create<RecipeStudioState>((set, get) => ({
         hf_token: "",
         hf_endpoint: "https://huggingface.co",
         local_file_name: "",
+        unstructured_upload_uid:
+          nextSourceType === "unstructured" ? makeUnstructuredUploadUid() : "",
         unstructured_file_ids: [],
         unstructured_file_names: [],
         unstructured_file_sizes: [],
@@ -446,7 +488,8 @@ export const useRecipeStudioStore = create<RecipeStudioState>((set, get) => ({
         activeConfigId: existing.id,
         dialogOpen: openDialog,
       };
-    }),
+    });
+  },
   addLlmNode: (type, position, openDialog = true) =>
     set((state) => {
       if (state.executionLocked) {
@@ -699,6 +742,9 @@ export const useRecipeStudioStore = create<RecipeStudioState>((set, get) => ({
       dialogOpen: false,
       sheetView: "root",
       fitViewTick: state.fitViewTick + 1,
+      // Queued cleanups belong to the previous recipe; draining them after
+      // a save of this one could delete files its saved payload still uses.
+      pendingUploadCleanups: [],
     })),
   setAuxNodePosition: (id, position) =>
     set((state) => {
@@ -786,6 +832,17 @@ export const useRecipeStudioStore = create<RecipeStudioState>((set, get) => ({
     set(applyUpdate);
   },
   onNodesChange: (changes) => {
+    const current = get();
+    if (!current.executionLocked) {
+      for (const change of changes) {
+        if (change.type === "remove") {
+          const uid = seedUploadCleanupUid(current.configs[change.id]);
+          if (uid) {
+            current.queueUploadCleanup(uid);
+          }
+        }
+      }
+    }
     const applyNodesChange = (state: RecipeStudioState) => {
       if (state.executionLocked) {
         return state;
diff --git a/studio/frontend/src/features/recipe-studio/types/index.ts b/studio/frontend/src/features/recipe-studio/types/index.ts
index b8ed13f70b..9231c5f7a3 100644
--- a/studio/frontend/src/features/recipe-studio/types/index.ts
+++ b/studio/frontend/src/features/recipe-studio/types/index.ts
@@ -340,6 +340,8 @@ export type SeedConfig = {
   hf_token?: string;
   hf_endpoint?: string;
   local_file_name?: string;
+  // ui-only: stable per-block id for uploads, since node ids collide across imports
+  unstructured_upload_uid?: string;
   unstructured_file_ids?: string[];
   unstructured_file_names?: string[];
   unstructured_file_sizes?: number[];
diff --git a/studio/frontend/src/features/recipe-studio/utils/config-factories.ts b/studio/frontend/src/features/recipe-studio/utils/config-factories.ts
index 76fc2c38ee..d47bac8858 100644
--- a/studio/frontend/src/features/recipe-studio/utils/config-factories.ts
+++ b/studio/frontend/src/features/recipe-studio/utils/config-factories.ts
@@ -20,6 +20,46 @@ import type {
 } from "../types";
 import { nextName } from "./naming";
 
+export function makeUnstructuredUploadUid(): string {
+  if (typeof globalThis.crypto?.randomUUID === "function") {
+    return globalThis.crypto.randomUUID().replace(/-/g, "").toLowerCase();
+  }
+  if (typeof globalThis.crypto?.getRandomValues === "function") {
+    const bytes = new Uint8Array(16);
+    globalThis.crypto.getRandomValues(bytes);
+    return Array.from(bytes, (byte) => byte.toString(16).padStart(2, "0")).join(
+      "",
+    );
+  }
+  let uid = "";
+  while (uid.length < 32) {
+    uid += Math.floor(Math.random() * 0x100000000)
+      .toString(16)
+      .padStart(8, "0");
+  }
+  return uid.slice(0, 32);
+}
+
+export function resolveUnstructuredUploadBlockId({
+  configId,
+  uploadUid,
+  generatedUploadUid,
+  unstructuredFileCount,
+}: {
+  configId: string;
+  uploadUid: string;
+  generatedUploadUid: string | null;
+  unstructuredFileCount: number;
+}): string {
+  if (uploadUid) {
+    return uploadUid;
+  }
+  if (generatedUploadUid) {
+    return generatedUploadUid;
+  }
+  return unstructuredFileCount > 0 ? configId : "";
+}
+
 export function makeSamplerConfig(
   id: string,
   samplerType: SamplerType,
@@ -368,6 +408,9 @@ export function makeSeedConfig(
     hf_token: "",
     hf_endpoint: "https://huggingface.co",
     local_file_name: "",
+    ...(seedSourceType === "unstructured"
+      ? { unstructured_upload_uid: makeUnstructuredUploadUid() }
+      : {}),
     unstructured_file_ids: [],
     unstructured_file_names: [],
     unstructured_file_sizes: [],
diff --git a/studio/frontend/src/features/recipe-studio/utils/import/importer.ts b/studio/frontend/src/features/recipe-studio/utils/import/importer.ts
index ac881d0373..54df2ffd50 100644
--- a/studio/frontend/src/features/recipe-studio/utils/import/importer.ts
+++ b/studio/frontend/src/features/recipe-studio/utils/import/importer.ts
@@ -16,11 +16,7 @@ import type {
 } from "../../types";
 import { buildEdges } from "./edges";
 import { isRecord, parseJson, readString } from "./helpers";
-import {
-  parseColumn,
-  parseModelConfig,
-  parseModelProvider,
-} from "./parsers";
+import { parseColumn, parseModelConfig, parseModelProvider } from "./parsers";
 import { parseSeedConfig } from "./parsers/seed-config-parser";
 import { buildNodes, parseUi } from "./ui";
 import type { ImportResult } from "./types";
@@ -43,6 +39,7 @@ type UiInput = {
   seed_drop_columns?: unknown;
   seed_preview_rows?: unknown;
   local_file_name?: unknown;
+  unstructured_upload_uid?: unknown;
   unstructured_file_ids?: unknown;
   unstructured_file_names?: unknown;
   unstructured_file_sizes?: unknown;
@@ -51,6 +48,10 @@ type UiInput = {
   advanced_open_by_node?: unknown;
 };
 
+type ImportRecipePayloadOptions = {
+  preserveUnstructuredUploads?: boolean;
+};
+
 type UiMarkdownNoteNode = {
   name: string;
   markdown: string;
@@ -90,7 +91,7 @@ function parseProcessors(input: unknown): RecipeProcessorConfig[] {
         ? templateRaw
         : isRecord(templateRaw)
           ? JSON.stringify(templateRaw, null, 2)
-          : "{\n  \"text\": \"{{ column_name }}\"\n}";
+          : '{\n  "text": "{{ column_name }}"\n}';
     processors.push({
       id: `p${index + 1}`,
       // biome-ignore lint/style/useNamingConvention: api schema
@@ -135,9 +136,7 @@ function parseSeedDropColumns(input: unknown): string[] {
   return Array.from(values);
 }
 
-function parseMcpProviders(
-  input: unknown,
-): Map<string, LlmMcpProviderConfig> {
+function parseMcpProviders(input: unknown): Map<string, LlmMcpProviderConfig> {
   const providers = new Map<string, LlmMcpProviderConfig>();
   if (!Array.isArray(input)) {
     return providers;
@@ -156,13 +155,12 @@ function parseMcpProviders(
     const args = Array.isArray(item.args)
       ? item.args.map((value) => String(value))
       : [];
-    const envPairs =
-      isRecord(item.env)
-        ? Object.entries(item.env).map(([key, value]) => ({
-            key: String(key),
-            value: String(value),
-          }))
-        : [];
+    const envPairs = isRecord(item.env)
+      ? Object.entries(item.env).map(([key, value]) => ({
+          key: String(key),
+          value: String(value),
+        }))
+      : [];
     providers.set(name, {
       id: `mcp-${index + 1}`,
       name,
@@ -209,7 +207,8 @@ function parseToolConfigs(input: unknown): Map<string, LlmToolConfig> {
       allow_tools: allowTools,
       // biome-ignore lint/style/useNamingConvention: api schema
       max_tool_call_turns:
-        item.max_tool_call_turns === null || item.max_tool_call_turns === undefined
+        item.max_tool_call_turns === null ||
+        item.max_tool_call_turns === undefined
           ? "5"
           : String(item.max_tool_call_turns),
       // biome-ignore lint/style/useNamingConvention: api schema
@@ -257,7 +256,9 @@ function parseUiMarkdownNoteNodes(input: unknown): UiMarkdownNoteNode[] {
   return noteNodes;
 }
 
-function parseUiToolProfileNodes(input: unknown): Map<string, Record<string, string[]>> {
+function parseUiToolProfileNodes(
+  input: unknown,
+): Map<string, Record<string, string[]>> {
   const toolProfiles = new Map<string, Record<string, string[]>>();
   if (!Array.isArray(input)) {
     return toolProfiles;
@@ -312,9 +313,15 @@ function parseAdvancedOpenByNode(input: unknown): Record<string, boolean> {
   return out;
 }
 
-type AdvancedOpenConfig = LlmConfig | SamplerConfig | SeedConfig | ValidatorConfig;
+type AdvancedOpenConfig =
+  | LlmConfig
+  | SamplerConfig
+  | SeedConfig
+  | ValidatorConfig;
 
-function isAdvancedOpenConfig(config: NodeConfig): config is AdvancedOpenConfig {
+function isAdvancedOpenConfig(
+  config: NodeConfig,
+): config is AdvancedOpenConfig {
   return (
     config.kind === "llm" ||
     config.kind === "sampler" ||
@@ -350,7 +357,8 @@ function buildToolProfileConfig(
       .map((providerName) => mcpProvidersByName.get(providerName))
       .flatMap((provider) => (provider ? [cloneMcpProvider(provider)] : [])),
     // biome-ignore lint/style/useNamingConvention: ui schema
-    fetched_tools_by_provider: fetchedToolsByProfileName.get(canonical.tool_alias) ?? {},
+    fetched_tools_by_provider:
+      fetchedToolsByProfileName.get(canonical.tool_alias) ?? {},
     // biome-ignore lint/style/useNamingConvention: api schema
     allow_tools: [...(canonical.allow_tools ?? [])],
     // biome-ignore lint/style/useNamingConvention: api schema
@@ -360,7 +368,10 @@ function buildToolProfileConfig(
   };
 }
 
-export function importRecipePayload(input: string): ImportResult {
+export function importRecipePayload(
+  input: string,
+  options: ImportRecipePayloadOptions = {},
+): ImportResult {
   const parsed = parseJson(input);
   if (!parsed.data || !isRecord(parsed.data)) {
     return {
@@ -369,9 +380,9 @@ export function importRecipePayload(input: string): ImportResult {
     };
   }
 
-  const recipe = (isRecord(parsed.data.recipe)
-    ? parsed.data.recipe
-    : parsed.data) as RecipeInput;
+  const recipe = (
+    isRecord(parsed.data.recipe) ? parsed.data.recipe : parsed.data
+  ) as RecipeInput;
   const ui = isRecord(parsed.data.ui) ? (parsed.data.ui as UiInput) : null;
 
   if (!Array.isArray(recipe.columns)) {
@@ -410,21 +421,36 @@ export function importRecipePayload(input: string): ImportResult {
         .map((row) => ({ ...row }))
     : undefined;
   const uiLocalFileName = readString(ui?.local_file_name) ?? undefined;
-  // Preserve file IDs/names from saved recipes (cleared at share time by sanitizeSeedForShare)
-  const uiUnstructuredFileIds: string[] = Array.isArray(ui?.unstructured_file_ids)
-    ? (ui.unstructured_file_ids as string[]).filter((v): v is string => typeof v === "string")
-    : [];
-  const uiUnstructuredFileNames: string[] = Array.isArray(ui?.unstructured_file_names)
-    ? (ui.unstructured_file_names as string[]).filter((v): v is string => typeof v === "string")
-    : [];
-  const uiUnstructuredFileSizes: number[] = Array.isArray(ui?.unstructured_file_sizes)
-    ? (ui.unstructured_file_sizes as number[]).filter((v): v is number => typeof v === "number")
-    : [];
+  const preserveUnstructuredUploads =
+    options.preserveUnstructuredUploads === true;
+  const uiUnstructuredUploadUid = preserveUnstructuredUploads
+    ? (readString(ui?.unstructured_upload_uid) ?? undefined)
+    : undefined;
+  const uiUnstructuredFileIds: string[] =
+    preserveUnstructuredUploads && Array.isArray(ui?.unstructured_file_ids)
+      ? (ui.unstructured_file_ids as string[]).filter(
+          (v): v is string => typeof v === "string",
+        )
+      : [];
+  const uiUnstructuredFileNames: string[] =
+    preserveUnstructuredUploads && Array.isArray(ui?.unstructured_file_names)
+      ? (ui.unstructured_file_names as string[]).filter(
+          (v): v is string => typeof v === "string",
+        )
+      : [];
+  const uiUnstructuredFileSizes: number[] =
+    preserveUnstructuredUploads && Array.isArray(ui?.unstructured_file_sizes)
+      ? (ui.unstructured_file_sizes as number[]).filter(
+          (v): v is number => typeof v === "number",
+        )
+      : [];
   const uiUnstructuredChunkSize = readStringNumber(ui?.unstructured_chunk_size);
   const uiUnstructuredChunkOverlap = readStringNumber(
     ui?.unstructured_chunk_overlap,
   );
-  const uiAdvancedOpenByNode = parseAdvancedOpenByNode(ui?.advanced_open_by_node);
+  const uiAdvancedOpenByNode = parseAdvancedOpenByNode(
+    ui?.advanced_open_by_node,
+  );
   const uiMarkdownNotes = parseUiMarkdownNoteNodes(ui?.nodes);
   const uiToolProfilesByName = parseUiToolProfileNodes(ui?.nodes);
 
@@ -459,11 +485,13 @@ export function importRecipePayload(input: string): ImportResult {
           : payloadSeedDropColumns,
       seed_preview_rows: uiSeedPreviewRows,
       local_file_name: uiLocalFileName,
+      unstructuredUploadUid: uiUnstructuredUploadUid,
       unstructuredFileIds: uiUnstructuredFileIds,
       unstructuredFileNames: uiUnstructuredFileNames,
       unstructuredFileSizes: uiUnstructuredFileSizes,
       unstructured_chunk_size: uiUnstructuredChunkSize,
       unstructured_chunk_overlap: uiUnstructuredChunkOverlap,
+      preserveUnstructuredUploads,
     });
     if (seedConfig) {
       applyAdvancedOpen(seedConfig, uiAdvancedOpenByNode);
@@ -567,12 +595,7 @@ export function importRecipePayload(input: string): ImportResult {
   const { layouts, auxNodes, edges: uiEdges, layoutDirection } = parseUi(ui);
   const resolvedLayoutDirection = layoutDirection ?? "LR";
   const nodes = buildNodes(configs, layouts);
-  const edges = buildEdges(
-    configs,
-    nameToId,
-    uiEdges,
-    resolvedLayoutDirection,
-  );
+  const edges = buildEdges(configs, nameToId, uiEdges, resolvedLayoutDirection);
   const auxNodePositions = Object.fromEntries(
     auxNodes.flatMap((item) => {
       const llmId = nameToId.get(item.llm);
@@ -583,10 +606,7 @@ export function importRecipePayload(input: string): ImportResult {
     }),
   );
 
-  const maxY = nodes.reduce(
-    (acc, node) => Math.max(acc, node.position.y),
-    0,
-  );
+  const maxY = nodes.reduce((acc, node) => Math.max(acc, node.position.y), 0);
 
   return {
     errors: [],
diff --git a/studio/frontend/src/features/recipe-studio/utils/import/parsers/seed-config-parser.ts b/studio/frontend/src/features/recipe-studio/utils/import/parsers/seed-config-parser.ts
index 21eadb3195..939205fe6d 100644
--- a/studio/frontend/src/features/recipe-studio/utils/import/parsers/seed-config-parser.ts
+++ b/studio/frontend/src/features/recipe-studio/utils/import/parsers/seed-config-parser.ts
@@ -197,17 +197,26 @@ export function parseSeedConfig(
     seed_drop_columns?: string[];
     seed_preview_rows?: Record<string, unknown>[];
     local_file_name?: string;
+    unstructuredUploadUid?: string;
     unstructuredFileIds?: string[];
     unstructuredFileNames?: string[];
     unstructuredFileSizes?: number[];
     unstructured_chunk_size?: string;
     unstructured_chunk_overlap?: string;
+    preserveUnstructuredUploads?: boolean;
   },
 ): SeedConfig | null {
   if (!seedConfigRaw) {
     return null;
   }
-  const parsed = parseSeedSettings(seedConfigRaw);
+  const parsed = { ...parseSeedSettings(seedConfigRaw) };
+  if (
+    parsed.seed_source_type === "unstructured" &&
+    options?.preserveUnstructuredUploads !== true
+  ) {
+    parsed.hf_path = "";
+    parsed.resolved_paths = [];
+  }
   let sourceType: SeedSourceType = "hf";
   if (parsed.seed_source_type === "hf") {
     sourceType = "hf";
@@ -230,6 +239,9 @@ export function parseSeedConfig(
     ...(options?.local_file_name !== undefined
       ? { local_file_name: options.local_file_name }
       : {}),
+    ...(options?.unstructuredUploadUid
+      ? { unstructured_upload_uid: options.unstructuredUploadUid }
+      : {}),
     ...(options?.unstructuredFileIds !== undefined
       ? { unstructured_file_ids: options.unstructuredFileIds }
       : {}),
diff --git a/studio/frontend/src/features/recipe-studio/utils/payload/build-payload.ts b/studio/frontend/src/features/recipe-studio/utils/payload/build-payload.ts
index 34c3c34274..9b2ad5b2b5 100644
--- a/studio/frontend/src/features/recipe-studio/utils/payload/build-payload.ts
+++ b/studio/frontend/src/features/recipe-studio/utils/payload/build-payload.ts
@@ -440,6 +440,9 @@ export function buildRecipePayload(
             unstructured_file_names: firstSeed.unstructured_file_names,
             unstructured_file_sizes: firstSeed.unstructured_file_sizes,
           }),
+        ...(firstSeed?.unstructured_upload_uid?.trim() && {
+          unstructured_upload_uid: firstSeed.unstructured_upload_uid,
+        }),
         ...(firstSeed &&
           firstSeed.unstructured_chunk_size !== undefined && {
             unstructured_chunk_size: firstSeed.unstructured_chunk_size,
diff --git a/studio/frontend/src/features/recipe-studio/utils/payload/types.ts b/studio/frontend/src/features/recipe-studio/utils/payload/types.ts
index 902ea796b4..763e68c1fb 100644
--- a/studio/frontend/src/features/recipe-studio/utils/payload/types.ts
+++ b/studio/frontend/src/features/recipe-studio/utils/payload/types.ts
@@ -71,6 +71,8 @@ export type RecipePayload = {
     seed_preview_rows?: Record<string, unknown>[];
     local_file_name?: string;
     // biome-ignore lint/style/useNamingConvention: api schema
+    unstructured_upload_uid?: string;
+    // biome-ignore lint/style/useNamingConvention: api schema
     unstructured_file_ids?: string[];
     // biome-ignore lint/style/useNamingConvention: api schema
     unstructured_file_names?: string[];
diff --git a/studio/frontend/src/features/settings/api/coding-agents.ts b/studio/frontend/src/features/settings/api/coding-agents.ts
new file mode 100644
index 0000000000..ae371b2d3a
--- /dev/null
+++ b/studio/frontend/src/features/settings/api/coding-agents.ts
@@ -0,0 +1,45 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { authFetch } from "@/features/auth";
+import { readFastApiError } from "@/lib/format-fastapi-error";
+
+export type CodingAgentsInfo = {
+  // Every agent `unsloth start` supports, in the CLI's declared order.
+  agents: string[];
+  // Subset of `agents` whose CLI binary was found on PATH by the backend.
+  detected: string[];
+};
+
+type ApiCodingAgentsInfo = {
+  agents: string[];
+  detected: string[];
+};
+
+// Which CLIs are on PATH is environment state, not a persisted setting -- it
+// can change any time the user installs something new, so this only
+// de-duplicates concurrent in-flight calls (e.g. React strict-mode's double
+// mount) rather than caching the result across the module's lifetime. Every
+// fresh call (each time a settings panel mounts) re-checks PATH for real.
+let inFlightInfo: Promise<CodingAgentsInfo> | null = null;
+
+function fromApi(info: ApiCodingAgentsInfo): CodingAgentsInfo {
+  return { agents: info.agents, detected: info.detected };
+}
+
+async function fetchCodingAgents(): Promise<CodingAgentsInfo> {
+  const res = await authFetch("/api/settings/coding-agents");
+  if (!res.ok) {
+    throw new Error(
+      await readFastApiError(res, "Failed to load installed coding agents"),
+    );
+  }
+  return fromApi(await res.json());
+}
+
+export async function loadCodingAgents(): Promise<CodingAgentsInfo> {
+  inFlightInfo ??= fetchCodingAgents().finally(() => {
+    inFlightInfo = null;
+  });
+  return inFlightInfo;
+}
diff --git a/studio/frontend/src/features/settings/components/agent-command.ts b/studio/frontend/src/features/settings/components/agent-command.ts
index 9e87922970..38b2d73c3b 100644
--- a/studio/frontend/src/features/settings/components/agent-command.ts
+++ b/studio/frontend/src/features/settings/components/agent-command.ts
@@ -12,7 +12,7 @@ const DEFAULT_AGENT = "claude";
 
 // URL.hostname brackets IPv6 literals (`new URL("http://[::1]:8888").hostname` is
 // "[::1]"), so strip the brackets before matching the bare "::1" loopback rules below.
-function normalizeHost(host: string): string {
+export function normalizeHost(host: string): string {
   const lower = host.toLowerCase();
   return lower.startsWith("[") && lower.endsWith("]") ? lower.slice(1, -1) : lower;
 }
@@ -26,7 +26,7 @@ function isDefaultLocalHost(host: string): boolean {
 }
 
 // Match the CLI auto-mint rule (is_loopback_url): localhost, ::1, and all of 127.0.0.0/8.
-function isLoopbackHost(host: string): boolean {
+export function isLoopbackHost(host: string): boolean {
   if (host === "localhost" || host === "::1") return true;
   const octets = host.split(".");
   return (
diff --git a/studio/frontend/src/features/settings/components/usage-examples.tsx b/studio/frontend/src/features/settings/components/usage-examples.tsx
index c43dd0f219..b3396c8d9b 100644
--- a/studio/frontend/src/features/settings/components/usage-examples.tsx
+++ b/studio/frontend/src/features/settings/components/usage-examples.tsx
@@ -16,6 +16,7 @@ import { fetchDeviceType, usePlatformStore } from "@/config/env";
 import { useChatRuntimeStore } from "@/features/chat";
 import { useT } from "@/i18n";
 import type { TranslationKey } from "@/i18n";
+import { isTauri } from "@/lib/api-base";
 import { copyToClipboard } from "@/lib/copy-to-clipboard";
 import { Tick02Icon } from "@/lib/tick-icon";
 import { cn } from "@/lib/utils";
@@ -25,14 +26,15 @@ import {
   InformationCircleIcon,
 } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
-import { useEffect, useMemo, useState } from "react";
+import { useEffect, useMemo, useRef, useState } from "react";
 import { Streamdown } from "streamdown";
+import { loadCodingAgents } from "../api/coding-agents";
 import {
   type OpenAIAutoSwitchSettings,
   loadOpenAIAutoSwitchSettings,
   updateOpenAIAutoSwitchSettings,
 } from "../api/openai-auto-switch";
-import { buildAgentCommand } from "./agent-command";
+import { buildAgentCommand, isLoopbackHost, normalizeHost } from "./agent-command";
 
 type ExampleType =
   | "curl"
@@ -114,6 +116,30 @@ const DOC_LINKS = [
   { label: "Hermes Agent", href: "https://unsloth.ai/docs/integrations/hermes-agent" },
 ];
 
+// Falls back to this list until the backend's installed-CLI check resolves;
+// kept in sync with the `unsloth start <agent>` subcommands and with
+// CODING_AGENTS in studio/backend/utils/coding_agents.py.
+const DEFAULT_AGENTS = [
+  "claude",
+  "codex",
+  "openclaw",
+  "opencode",
+  "hermes",
+  "pi",
+];
+// The agent selection resets to this whenever an auto-pick is no longer
+// trustworthy (leaving loopback, or the only compatible detected agent
+// stops being compatible) rather than lingering on a stale choice.
+const DEFAULT_AGENT = "claude";
+const AGENT_LABELS: Record<string, string> = {
+  claude: "Claude Code",
+  codex: "Codex",
+  openclaw: "OpenClaw",
+  opencode: "OpenCode",
+  hermes: "Hermes",
+  pi: "Pi",
+};
+
 const j = (s: string): string => JSON.stringify(s);
 const shSingle = (s: string): string => s.replace(/'/g, "'\\''");
 const psSingle = (s: string): string => s.replace(/'/g, "''");
@@ -399,6 +425,17 @@ function useLoadedModelName(): string {
   }, [checkpoint, ggufVariant]);
 }
 
+// Backend PATH detection is only safe in the desktop app, where the UI owns
+// the local backend. A browser loopback URL may be an SSH/local port forward.
+function canUseLocalAgentDetection(base: string): boolean {
+  if (!isTauri) return false;
+  try {
+    return isLoopbackHost(normalizeHost(new URL(base).hostname));
+  } catch {
+    return false;
+  }
+}
+
 const SHIKI_THEMES = [unslothLightTheme, unslothDarkTheme] as [
   typeof unslothLightTheme,
   typeof unslothDarkTheme,
@@ -443,7 +480,18 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
   const [copied, setCopied] = useState(false);
   const [copiedUrl, setCopiedUrl] = useState(false);
   const [copiedAgent, setCopiedAgent] = useState(false);
+  const [agent, setAgent] = useState<string>(DEFAULT_AGENT);
+  const [availableAgents, setAvailableAgents] =
+    useState<string[]>(DEFAULT_AGENTS);
+  const [detectedAgents, setDetectedAgents] = useState<string[]>([]);
+  // True once the user has picked an agent themselves; guards the detection
+  // effect below from clobbering that choice if it resolves afterward.
+  const agentPickedByUserRef = useRef(false);
   const [useTunnel, setUseTunnel] = useState<boolean>(readUseTunnelPref);
+  const origin = typeof window !== "undefined" ? window.location.origin : "";
+  const base =
+    useTunnel && cloudflareUrl ? cloudflareUrl : (serverUrl ?? origin);
+  const localAgentDetection = canUseLocalAgentDetection(base);
   // null while loading; the same setting the General tab exposes (shared cache).
   const [autoSwitch, setAutoSwitch] = useState<OpenAIAutoSwitchSettings | null>(
     null,
@@ -454,6 +502,78 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
     void fetchDeviceType({ force: true });
   }, []);
 
+  // Fetching is the only job of this effect: populate availableAgents/
+  // detectedAgents (or clear them). Which agent gets auto-picked from that
+  // list is derived separately below, so it can react to the loaded model
+  // changing too, not just a fresh fetch.
+  useEffect(() => {
+    // Browser loopback URLs can be SSH/local forwards, so only the desktop app
+    // may use backend PATH checks to mark or auto-pick local agents.
+    if (!localAgentDetection) {
+      setDetectedAgents([]);
+      // A previously auto-picked agent was only ever verified against the
+      // Studio backend's PATH, which is meaningless now that this panel no
+      // longer targets a loopback base -- don't leave it selected, but
+      // never touch a choice the user made by hand.
+      if (!agentPickedByUserRef.current) {
+        setAgent(DEFAULT_AGENT);
+      }
+      return;
+    }
+
+    let cancelled = false;
+    void loadCodingAgents()
+      .then((info) => {
+        if (cancelled) return;
+        setAvailableAgents(info.agents);
+        setDetectedAgents(info.detected);
+      })
+      .catch(() => {
+        // Best-effort: keep the default agent list and let the user pick manually.
+      });
+    return () => {
+      cancelled = true;
+    };
+  }, [localAgentDetection]);
+
+  // Single source of truth for the auto-picked agent, re-derived whenever
+  // the detected list or the loaded model's GGUF-ness changes -- in either
+  // direction. `codex` needs a GGUF model (unsloth_cli's
+  // _require_gguf_for_codex exits otherwise), so it's only preferred once
+  // the loaded model actually qualifies; loading a GGUF model *after* a
+  // non-GGUF-gated fallback picked something else re-steers back to codex
+  // just as loading a non-GGUF model steers away from it. Never overrides a
+  // choice the user made by hand.
+  // activeGgufVariant alone only covers an HF-repo GGUF pick (a specific
+  // quant variant string) -- a direct local .gguf file (custom folder /
+  // LM Studio / drag-drop) is just as much a GGUF the codex preflight would
+  // accept, but never has a "variant" to report, and would otherwise read as
+  // non-GGUF here. activeNativePathToken covers the drag-drop/picked-file
+  // case; ggufContextLength is only ever populated when the backend's
+  // /api/inference/status last reported is_gguf: true for the active model
+  // (see applyActiveModelStatusToStore), so together these three cover every
+  // path a model can be GGUF through, matching the same is_gguf-or-equivalent
+  // check hasGgufSource applies to a staged pick.
+  const activeGgufVariant = useChatRuntimeStore((s) => s.activeGgufVariant);
+  const activeNativePathToken = useChatRuntimeStore((s) => s.activeNativePathToken);
+  const ggufContextLength = useChatRuntimeStore((s) => s.ggufContextLength);
+  useEffect(() => {
+    if (agentPickedByUserRef.current) return;
+    if (detectedAgents.length === 0) return;
+    const isGguf =
+      activeGgufVariant != null || activeNativePathToken != null || ggufContextLength != null;
+    const preferred = detectedAgents.find((a) => a !== "codex" || isGguf);
+    if (preferred) {
+      setAgent(preferred);
+    } else if (agent === "codex" && !isGguf) {
+      // codex was auto-picked while a GGUF model was active and it's the
+      // only detected agent; now that the model isn't GGUF anymore, nothing
+      // detected is actually runnable, so fall back to the default instead
+      // of leaving a codex command unsloth_cli will reject.
+      setAgent(DEFAULT_AGENT);
+    }
+  }, [agent, detectedAgents, activeGgufVariant, activeNativePathToken, ggufContextLength]);
+
   useEffect(() => {
     let cancelled = false;
     void loadOpenAIAutoSwitchSettings()
@@ -470,9 +590,6 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
 
   const model = useLoadedModelName();
   const key = apiKey || KEY_PLACEHOLDER;
-  const origin = typeof window !== "undefined" ? window.location.origin : "";
-  const base =
-    useTunnel && cloudflareUrl ? cloudflareUrl : (serverUrl ?? origin);
 
   const autoSwitchOn = autoSwitch?.enabled ?? false;
   const snippets = useMemo(
@@ -481,8 +598,8 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
   );
   // Agent command must target the server the panel shows, not the :8888 default.
   const agentCommand = useMemo(
-    () => buildAgentCommand(base, key, os),
-    [base, key, os],
+    () => buildAgentCommand(base, key, os, agent),
+    [base, key, os, agent],
   );
 
   const osAware = OS_AWARE[lang];
@@ -710,6 +827,42 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
           <span className="text-[11px] leading-snug text-muted-foreground">
             {t("settings.apiKeys.codingAgentsHint")}
           </span>
+          <div className="flex min-w-0 flex-wrap items-center gap-1">
+            {availableAgents.map((id) => {
+              const installed = detectedAgents.includes(id);
+              const active = agent === id;
+              return (
+                <button
+                  key={id}
+                  type="button"
+                  onClick={() => {
+                    agentPickedByUserRef.current = true;
+                    setAgent(id);
+                  }}
+                  aria-pressed={active}
+                  title={
+                    installed
+                      ? t("settings.apiKeys.codingAgentDetected")
+                      : undefined
+                  }
+                  className={cn(
+                    "flex items-center gap-1 rounded-full px-2.5 py-1 text-[11px] font-medium transition-colors focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring",
+                    active
+                      ? "hub-tab-toggle-pill text-foreground"
+                      : "text-muted-foreground hover:text-foreground",
+                  )}
+                >
+                  {AGENT_LABELS[id] ?? id}
+                  {installed ? (
+                    <span
+                      aria-hidden="true"
+                      className="size-1.5 rounded-full bg-emerald-500"
+                    />
+                  ) : null}
+                </button>
+              );
+            })}
+          </div>
           <div className="relative mt-0.5 min-w-0">
             <code className="block min-w-0 overflow-x-auto rounded border border-border bg-muted/30 px-2 py-1.5 pr-14 font-mono text-[11px] text-foreground">
               {agentCommand}
@@ -727,7 +880,13 @@ export function UsageExamples({ apiKey }: { apiKey?: string | null }) {
             </button>
           </div>
           <span className="text-[11px] leading-snug text-muted-foreground">
-            {t("settings.apiKeys.codingAgentsSwap")}
+            {detectedAgents.length > 0
+              ? t("settings.apiKeys.codingAgentsDetectedHint", {
+                  agents: detectedAgents
+                    .map((id) => AGENT_LABELS[id] ?? id)
+                    .join(", "),
+                })
+              : t("settings.apiKeys.codingAgentsSwap")}
           </span>
         </div>
         <div className="flex flex-wrap items-center gap-x-2 gap-y-1 border-t border-border px-3 py-2 text-[11px] text-muted-foreground">
diff --git a/studio/frontend/src/features/settings/index.ts b/studio/frontend/src/features/settings/index.ts
index 364ca1611f..3fefd8c63a 100644
--- a/studio/frontend/src/features/settings/index.ts
+++ b/studio/frontend/src/features/settings/index.ts
@@ -7,6 +7,7 @@ export {
   savePersonalization,
 } from "./api/personalization";
 export { setTheme, useTheme } from "./stores/theme-store";
+export { useMonitorOverlayStore } from "./stores/monitor-overlay-store";
 export type {
   Personalization,
   PersonalizationAppearance,
diff --git a/studio/frontend/src/features/settings/settings-dialog.tsx b/studio/frontend/src/features/settings/settings-dialog.tsx
index 0d201edc0d..d8a0d45d3f 100644
--- a/studio/frontend/src/features/settings/settings-dialog.tsx
+++ b/studio/frontend/src/features/settings/settings-dialog.tsx
@@ -56,6 +56,7 @@ const TABS: TabDef[] = [
     id: "resources",
     labelKey: "settings.tabs.resources",
     icon: CpuIcon,
+    badgeKey: "common.new",
   },
   {
     id: "chat",
@@ -72,7 +73,6 @@ const TABS: TabDef[] = [
     id: "connections",
     labelKey: "settings.tabs.connections",
     icon: CloudIcon,
-    badgeKey: "common.new",
   },
   { id: "about", labelKey: "settings.tabs.about", icon: HelpCircleIcon },
 ];
diff --git a/studio/frontend/src/i18n/locales/en.ts b/studio/frontend/src/i18n/locales/en.ts
index b67fd5ca1d..e0cb8030ae 100644
--- a/studio/frontend/src/i18n/locales/en.ts
+++ b/studio/frontend/src/i18n/locales/en.ts
@@ -409,7 +409,6 @@ export const en = {
       description: "Access Unsloth via the OpenAI-compatible API.",
       readDocs: "Read the API docs",
       noAccess: "No API access yet.",
-      newBadge: "New",
       accessTokens: "Access tokens",
       loadError: "Couldn't load API access.",
       createError: "Couldn't create access token.",
@@ -445,6 +444,8 @@ export const en = {
       codingAgentsHint:
         "Launch a coding agent against this server. It uses the loaded model; a local server mints an API key automatically, a remote one includes it in the command.",
       codingAgentsSwap: "Swap claude for codex, openclaw, opencode, hermes, or pi.",
+      codingAgentDetected: "Installed on this machine",
+      codingAgentsDetectedHint: "Detected on this machine: {agents}.",
       relativeNever: "never",
       relativeJustNow: "just now",
       relativeHoursAgo: "{count}h ago",
diff --git a/studio/frontend/src/i18n/locales/ja.ts b/studio/frontend/src/i18n/locales/ja.ts
index f2020f3cfb..08b8d4f4d3 100644
--- a/studio/frontend/src/i18n/locales/ja.ts
+++ b/studio/frontend/src/i18n/locales/ja.ts
@@ -296,7 +296,6 @@ export const ja = {
       description: "OpenAI互換 API を介して Unsloth にアクセスします。",
       readDocs: "API ドキュメントを読む",
       noAccess: "まだ API アクセス権がありません。",
-      newBadge: "新規",
       accessTokens: "アクセストークン",
       loadError: "API アクセス権を読み込めませんでした。",
       createError: "アクセストークンを作成できませんでした。",
diff --git a/studio/frontend/src/i18n/locales/pt-br.ts b/studio/frontend/src/i18n/locales/pt-br.ts
index c261e4ed0c..494a98ec32 100644
--- a/studio/frontend/src/i18n/locales/pt-br.ts
+++ b/studio/frontend/src/i18n/locales/pt-br.ts
@@ -363,7 +363,6 @@ export const ptBR = {
         "Acesse o Unsloth por meio da API compatível com OpenAI.",
       readDocs: "Leia a documentação da API",
       noAccess: "Nenhum acesso à API ainda.",
-      newBadge: "Novo",
       accessTokens: "Tokens de acesso",
       loadError: "Não foi possível carregar o acesso à API.",
       createError: "Não foi possível criar o token de acesso.",
diff --git a/studio/frontend/src/i18n/locales/zh-CN.ts b/studio/frontend/src/i18n/locales/zh-CN.ts
index f6dc265fc7..dda8e017ab 100644
--- a/studio/frontend/src/i18n/locales/zh-CN.ts
+++ b/studio/frontend/src/i18n/locales/zh-CN.ts
@@ -267,7 +267,6 @@ export const zhCN = {
       description: "通过兼容 OpenAI 的 API 以编程方式访问 Unsloth。",
       readDocs: "阅读 API 文档",
       noAccess: "还没有 API 访问权限。",
-      newBadge: "新",
       accessTokens: "访问 token",
       loadError: "无法加载 API 访问权限。",
       createError: "无法创建访问 token。",
diff --git a/studio/frontend/src/lib/latex.ts b/studio/frontend/src/lib/latex.ts
index 86a9634048..edf9875602 100644
--- a/studio/frontend/src/lib/latex.ts
+++ b/studio/frontend/src/lib/latex.ts
@@ -173,7 +173,11 @@ function looksLikeMathBody(body: string): boolean {
  * (`**$X$**`, `__$X$__`) are always math: LLMs use that for "bold math"
  * and the heuristic would otherwise reject prose-shaped bodies like "90 - x".
  */
-function hasInlineMathCloser(content: string, offset: number): boolean {
+function hasInlineMathCloser(
+  content: string,
+  offset: number,
+  mathRegions: Array<[number, number]>,
+): boolean {
   const MAX_SPAN = 200;
   const limit = Math.min(content.length, offset + 1 + MAX_SPAN);
   for (let i = offset + 1; i < limit; i++) {
@@ -181,6 +185,9 @@ function hasInlineMathCloser(content: string, offset: number): boolean {
     if (c === "\n") return false;
     if (c !== "$") continue;
     if (content[i - 1] === "\\") continue;
+    // A `$` opening a generated span (from `\(...\)`) is not a currency closer;
+    // pairing with it would swallow the price into math (`$5 + x \(y\)`).
+    if (isInRegion(i, mathRegions)) return false;
     if (content[i + 1] === "$") {
       i++;
       continue;
@@ -294,7 +301,22 @@ function convertLatexDelimiters(content: string): {
       continue;
     }
     append(content.slice(last, match.index));
-    const wrapped = isDisplay ? `\n$$\n${body}\n$$\n` : `$${body}$`;
+    let wrapped: string;
+    if (isDisplay) {
+      // Keep the opener's leading indentation so a `$$` block inside a list item
+      // stays in the container instead of breaking out at column 0. Only when the
+      // opener is whitespace-prefixed, so inline `text \[x\]` keeps column 0.
+      const lineStart =
+        match.index > 0 ? content.lastIndexOf("\n", match.index - 1) + 1 : 0;
+      const prefix = content.slice(lineStart, match.index);
+      const indent = /^\s*$/.test(prefix) ? prefix : "";
+      // Indent every body line, not just the first, so multi-line display math
+      // (`\[a\nb\]`) stays wholly inside the container.
+      const inner = indent ? body.replace(/\n/g, `\n${indent}`) : body;
+      wrapped = `\n${indent}$$\n${indent}${inner}\n${indent}$$\n`;
+    } else {
+      wrapped = `$${body}$`;
+    }
     const start = append(wrapped);
     mathRegions.push([start, offset]);
     last = matchEnd;
@@ -334,7 +356,7 @@ export function preprocessLaTeX(content: string): string {
     if (isInRegion(offset, mathRegions)) {
       return match;
     }
-    if (hasInlineMathCloser(text, offset)) {
+    if (hasInlineMathCloser(text, offset, mathRegions)) {
       return match;
     }
     return "\\" + match;
diff --git a/studio/install_llama_prebuilt.py b/studio/install_llama_prebuilt.py
index e40cb3083e..ca1fe79efa 100644
--- a/studio/install_llama_prebuilt.py
+++ b/studio/install_llama_prebuilt.py
@@ -10,6 +10,7 @@ import argparse
 import atexit
 import errno
 import fnmatch
+import glob
 import hashlib
 import json
 import os
@@ -165,9 +166,9 @@ def env_int(
 # errors. Only use "master" temporarily when the latest release is missing
 # support for a new model architecture.
 DEFAULT_LLAMA_TAG = os.environ.get("UNSLOTH_LLAMA_TAG", "latest")
-# Default published repo for prebuilt release resolution. Linux uses
-# Unsloth prebuilts; setup.sh/setup.ps1 pass --published-repo explicitly
-# for macOS/Windows to override with ggml-org/llama.cpp when needed.
+# Default published repo for prebuilt release resolution. Every host plans
+# its prebuilt against the Unsloth fork; setup.sh/setup.ps1 pass it via
+# --published-repo. ggml-org is reachable only via an explicit override.
 DEFAULT_PUBLISHED_REPO = "unslothai/llama.cpp"
 DEFAULT_PUBLISHED_TAG = os.environ.get("UNSLOTH_LLAMA_RELEASE_TAG")
 DEFAULT_PUBLISHED_MANIFEST_ASSET = os.environ.get(
@@ -265,6 +266,7 @@ class HostInfo:
     has_physical_nvidia: bool
     has_usable_nvidia: bool
     has_rocm: bool = False
+    has_intel_gpu: bool = False
     rocm_gfx_target: str | None = None
     # (major, minor) from platform.mac_ver(); None off macOS or if unparseable.
     # Skips a macos prebuilt whose minimum-OS exceeds this host.
@@ -1284,162 +1286,6 @@ def synthetic_checksums_for_release(
     )
 
 
-def parse_direct_linux_release_bundle(
-    repo: str, release: dict[str, Any]
-) -> PublishedReleaseBundle | None:
-    release_tag = release.get("tag_name")
-    if not isinstance(release_tag, str) or not release_tag:
-        return None
-
-    assets = release_asset_map(release)
-    artifacts: list[PublishedLlamaArtifact] = []
-    inferred_labels: list[str] = []
-
-    linux_asset_re = re.compile(
-        r"^app-(?P<label>.+)-(?P<target>linux-x64(?:-cpu)?|linux-x64-cuda\d+-(?:older|newer|portable))\.tar\.gz$"
-    )
-    for asset_name in sorted(assets):
-        match = linux_asset_re.fullmatch(asset_name)
-        if not match:
-            continue
-        inferred_labels.append(match.group("label"))
-        target = match.group("target")
-        if target in {"linux-x64", "linux-x64-cpu"}:
-            artifacts.append(
-                PublishedLlamaArtifact(
-                    asset_name = asset_name,
-                    install_kind = "linux-cpu",
-                    runtime_line = None,
-                    coverage_class = None,
-                    supported_sms = [],
-                    min_sm = None,
-                    max_sm = None,
-                    bundle_profile = None,
-                    rank = 1000,
-                )
-            )
-            continue
-
-        bundle_profile = target.removeprefix("linux-x64-")
-        profile = _resolve_linux_bundle_profile(bundle_profile)
-        if profile is None:
-            continue
-        artifacts.append(
-            PublishedLlamaArtifact(
-                asset_name = asset_name,
-                install_kind = "linux-cuda",
-                runtime_line = str(profile["runtime_line"]),
-                coverage_class = str(profile["coverage_class"]),
-                supported_sms = [str(value) for value in profile["supported_sms"]],
-                min_sm = int(profile["min_sm"]),
-                max_sm = int(profile["max_sm"]),
-                bundle_profile = bundle_profile,
-                rank = int(profile["rank"]),
-            )
-        )
-
-    if not artifacts:
-        return None
-
-    upstream_tag = (
-        release_tag
-        if is_release_tag_like(release_tag)
-        else inferred_labels[0]
-        if len(set(inferred_labels)) == 1 and inferred_labels
-        else release_tag
-    )
-    selection_log = [
-        f"published_release: repo={repo}",
-        f"published_release: tag={release_tag}",
-        f"published_release: upstream_tag={upstream_tag}",
-        "published_release: direct_asset_scan=linux",
-    ]
-    return PublishedReleaseBundle(
-        repo = repo,
-        release_tag = release_tag,
-        upstream_tag = upstream_tag,
-        assets = assets,
-        manifest_asset_name = DEFAULT_PUBLISHED_MANIFEST_ASSET,
-        artifacts = artifacts,
-        selection_log = selection_log,
-    )
-
-
-def direct_linux_release_plan(
-    release: dict[str, Any], host: HostInfo, repo: str, requested_tag: str
-) -> InstallReleasePlan | None:
-    bundle = parse_direct_linux_release_bundle(repo, release)
-    if bundle is None:
-        return None
-    if not direct_release_matches_request(
-        release_tag = bundle.release_tag,
-        llama_tag = bundle.upstream_tag,
-        requested_tag = requested_tag,
-    ):
-        return None
-
-    attempts: list[AssetChoice] = []
-    if host.has_usable_nvidia:
-        # Prefer the cudart major Studio loads at runtime (torch's bundled
-        # libcudart), not the newest on disk. Otherwise a stray cuda13
-        # runtime outranks the torch cuda12 the binary links against.
-        torch_preference = detect_torch_cuda_runtime_preference(host)
-        selection = linux_cuda_choice_from_release(
-            host,
-            bundle,
-            preferred_runtime_line = torch_preference.runtime_line,
-            selection_preamble = torch_preference.selection_log,
-        )
-        if selection is not None:
-            attempts.extend(selection.attempts)
-    elif not host.has_rocm:
-        # A ROCm-only host gets no CPU asset: leaving attempts empty lets the
-        # raise below trigger a HIP source build instead of shipping a CPU
-        # binary on a GPU host (this ggml-org path has no per-gfx ROCm asset).
-        cpu_choice = published_asset_choice_for_kind(bundle, "linux-cpu")
-        if cpu_choice is not None:
-            attempts.append(cpu_choice)
-    # NVIDIA hosts whose CUDA selection produced nothing fall through to the
-    # raise below (mirroring the ROCm policy above): the caller then walks
-    # back to an older release that still ships a usable CUDA line instead of
-    # silently installing a CPU binary on a GPU host. Today's walk-back only
-    # works because partial releases ship no CPU bundle; this keeps it working
-    # if a future partial release does.
-    if not attempts:
-        raise PrebuiltFallback("no compatible Linux prebuilt asset was found")
-    approved_checksums = synthetic_checksums_for_release(
-        repo,
-        bundle.release_tag,
-        bundle.upstream_tag,
-    )
-    resolved_upstream_tag = bundle.upstream_tag
-    if DEFAULT_PUBLISHED_SHA256_ASSET in bundle.assets and not is_release_tag_like(
-        bundle.upstream_tag
-    ):
-        approved_checksums = load_approved_release_checksums(repo, bundle.release_tag)
-        # Require exact source provenance for branch/pull/commit releases.
-        # Mirrors validated_checksums_for_bundle so incomplete metadata fails
-        # closed instead of degrading to the legacy branch-as-tag source
-        # hydration path this PR eliminates.
-        if (
-            not approved_checksums.source_commit
-            or exact_source_archive_hash(approved_checksums) is None
-            or source_clone_url_from_checksums(approved_checksums) is None
-        ):
-            raise PrebuiltFallback(
-                f"approved checksum asset {DEFAULT_PUBLISHED_SHA256_ASSET} for "
-                f"{repo}@{bundle.release_tag} did not contain exact source provenance"
-            )
-        attempts = apply_approved_hashes(attempts, approved_checksums)
-    return InstallReleasePlan(
-        requested_tag = requested_tag,
-        llama_tag = resolved_upstream_tag,
-        release_tag = bundle.release_tag,
-        attempts = attempts,
-        approved_checksums = approved_checksums,
-    )
-
-
 def direct_upstream_release_plan(
     release: dict[str, Any], host: HostInfo, repo: str, requested_tag: str
 ) -> InstallReleasePlan | None:
@@ -1482,6 +1328,24 @@ def direct_upstream_release_plan(
                         install_kind = "windows-hip",
                     )
                 )
+        # Intel (or other non-NVIDIA/non-AMD) GPU: use the Vulkan prebuilt. Gate
+        # on no PHYSICAL NVIDIA (not just no usable one): a host that hid NVIDIA
+        # via CUDA_VISIBLE_DEVICES must not reach Vulkan, which ignores that mask
+        # and could enumerate the reserved card. Falls through to CPU below.
+        elif host.has_intel_gpu and not host.has_physical_nvidia:
+            vulkan_asset = f"llama-{release_tag}-bin-win-vulkan-x64.zip"
+            vulkan_url = assets.get(vulkan_asset)
+            if vulkan_url:
+                attempts.append(
+                    AssetChoice(
+                        repo = repo,
+                        tag = release_tag,
+                        name = vulkan_asset,
+                        url = vulkan_url,
+                        source_label = "upstream",
+                        install_kind = "windows-vulkan",
+                    )
+                )
         cpu_asset = f"llama-{release_tag}-bin-win-cpu-x64.zip"
         cpu_url = assets.get(cpu_asset)
         if cpu_url:
@@ -1545,6 +1409,23 @@ def direct_upstream_release_plan(
         # ROCm hosts are excluded: this ggml-org path ships no per-gfx ROCm
         # asset, so they fall through to the empty-attempts raise (HIP source
         # build) rather than silently getting a CPU binary on a GPU host.
+        # Intel (or other non-NVIDIA/non-AMD) GPU: use the Vulkan prebuilt. The
+        # elif already excludes usable NVIDIA and ROCm; also require no PHYSICAL
+        # NVIDIA so a CUDA-hidden card isn't reached through Vulkan (CPU below).
+        if host.has_intel_gpu and not host.has_physical_nvidia:
+            vulkan_asset = f"llama-{release_tag}-bin-ubuntu-vulkan-x64.tar.gz"
+            vulkan_url = assets.get(vulkan_asset)
+            if vulkan_url:
+                attempts.append(
+                    AssetChoice(
+                        repo = repo,
+                        tag = release_tag,
+                        name = vulkan_asset,
+                        url = vulkan_url,
+                        source_label = "upstream",
+                        install_kind = "linux-vulkan",
+                    )
+                )
         asset_name = f"llama-{release_tag}-bin-ubuntu-x64.tar.gz"
         asset_url = assets.get(asset_name)
         if asset_url:
@@ -1564,6 +1445,23 @@ def direct_upstream_release_plan(
         # selector returned 0 attempts and the installer fell back to a
         # source build on every Linux ARM64 host (DGX Spark, Ampere
         # Altra, GitHub-hosted ubuntu-24.04-arm runners, etc.).
+        # Intel (or other non-NVIDIA/non-AMD) GPU: prefer the Vulkan prebuilt,
+        # mirroring the x86_64 branch. Upstream ships bin-ubuntu-vulkan-arm64.
+        # No physical NVIDIA: don't reach a CUDA-hidden card through Vulkan.
+        if host.has_intel_gpu and not host.has_physical_nvidia and not host.has_rocm:
+            vulkan_asset = f"llama-{release_tag}-bin-ubuntu-vulkan-arm64.tar.gz"
+            vulkan_url = assets.get(vulkan_asset)
+            if vulkan_url:
+                attempts.append(
+                    AssetChoice(
+                        repo = repo,
+                        tag = release_tag,
+                        name = vulkan_asset,
+                        url = vulkan_url,
+                        source_label = "upstream",
+                        install_kind = "linux-vulkan",
+                    )
+                )
         asset_name = f"llama-{release_tag}-bin-ubuntu-arm64.tar.gz"
         asset_url = assets.get(asset_name)
         if asset_url:
@@ -2864,6 +2762,64 @@ def _pick_rocm_gfx_target(out: str) -> str | None:
     return _tokens[0]
 
 
+# Display-adapter device class: one NNNN subkey per installed display driver
+# config, each carrying the driver's DriverDesc and PCI MatchingDeviceId.
+_WINDOWS_DISPLAY_CLASS_KEY = (
+    r"SYSTEM\CurrentControlSet\Control\Class\{4d36e968-e325-11ce-bfc1-08002be10318}"
+)
+
+
+def windows_intel_gpu_in_registry() -> bool:
+    """Whether the Windows registry lists an Intel display adapter.
+
+    In-process Windows counterpart of the Linux DRM vendor-id check (0x8086),
+    with weaker semantics: the class key lists installed display-driver
+    configs, which can outlive removed hardware, where sysfs lists present
+    devices. A stale Intel entry at worst routes to the upstream Vulkan
+    prebuilt instead of the fork CPU bundle: inference still works (the
+    Vulkan build runs on CPU when no Vulkan device exists), at the cost of
+    fork-only extras such as the DiffusionGemma visual server. detect_host's
+    PowerShell + WMI probe can silently miss a real Intel GPU: a cold
+    powershell.exe start plus the first CIM query routinely exceeds the 15s
+    budget on hosts with slow AV scanning or a degraded WMI repository, and
+    the probe swallows the timeout (#4452, Arc A770 routed to the CPU
+    prebuilt). Reading the display-adapter class key needs no subprocess and
+    answers in microseconds. Matches the PCI vendor id in MatchingDeviceId
+    (ven_8086) or an Intel DriverDesc.
+    """
+    try:
+        import winreg
+    except ImportError:
+        return False
+    try:
+        with winreg.OpenKey(winreg.HKEY_LOCAL_MACHINE, _WINDOWS_DISPLAY_CLASS_KEY) as class_key:
+            for index in range(winreg.QueryInfoKey(class_key)[0]):
+                try:
+                    name = winreg.EnumKey(class_key, index)
+                    if not name.isdigit():
+                        # "Properties" is ACL-restricted and not an adapter.
+                        continue
+                    with winreg.OpenKey(class_key, name) as adapter_key:
+                        for value_name, needle in (
+                            ("MatchingDeviceId", "ven_8086"),
+                            ("DriverDesc", "intel"),
+                        ):
+                            try:
+                                value, _ = winreg.QueryValueEx(adapter_key, value_name)
+                            except OSError:
+                                continue
+                            if needle in str(value).lower():
+                                return True
+                except OSError:
+                    continue
+    except Exception:
+        # Advisory probe: any unexpected failure must degrade to the CIM
+        # fallback, never crash the installer (mirrors detect_host's own
+        # swallow around the CIM probe).
+        return False
+    return False
+
+
 def detect_host() -> HostInfo:
     system = platform.system()
     machine = platform.machine().lower()
@@ -3075,6 +3031,46 @@ def detect_host() -> HostInfo:
         # Note: amdhip64.dll presence alone is NOT treated as GPU evidence
         # since the HIP SDK can be installed without an AMD GPU.
 
+    # Detect an Intel GPU; gates the Vulkan prebuilt. Linux reads the DRM sysfs
+    # vendor id (0x8086); Windows reads the display-adapter registry class,
+    # then falls back to the WMI video controller list. Only probed with no
+    # usable NVIDIA and no ROCm (matching the Vulkan branches), keeping the
+    # probe (notably the Windows powershell call) off that path.
+    has_intel_gpu = False
+    if not has_usable_nvidia and not has_rocm:
+        if is_linux:
+            for _vendor_file in glob.glob("/sys/class/drm/card*/device/vendor"):
+                try:
+                    with open(_vendor_file) as _vf:
+                        if _vf.read().strip().lower() == "0x8086":
+                            has_intel_gpu = True
+                            break
+                except OSError:
+                    continue
+        elif is_windows:
+            # Registry first (in-process; see windows_intel_gpu_in_registry).
+            # The CIM query stays as the fallback when the registry shows no
+            # Intel adapter.
+            has_intel_gpu = windows_intel_gpu_in_registry()
+            if not has_intel_gpu:
+                _ps = shutil.which("powershell") or shutil.which("pwsh")
+                if _ps:
+                    try:
+                        _result = run_capture(
+                            [
+                                _ps,
+                                "-NoProfile",
+                                "-Command",
+                                "Get-CimInstance Win32_VideoController | "
+                                "Select-Object -ExpandProperty Name",
+                            ],
+                            timeout = 15,
+                        )
+                        if _result.returncode == 0 and "intel" in _result.stdout.lower():
+                            has_intel_gpu = True
+                    except Exception:
+                        pass
+
     return HostInfo(
         system = system,
         machine = machine,
@@ -3090,6 +3086,7 @@ def detect_host() -> HostInfo:
         has_physical_nvidia = has_physical_nvidia,
         has_usable_nvidia = has_usable_nvidia,
         has_rocm = has_rocm,
+        has_intel_gpu = has_intel_gpu,
         rocm_gfx_target = rocm_gfx_target,
         macos_version = macos_version,
     )
@@ -3126,6 +3123,7 @@ def _apply_host_overrides(
             has_physical_nvidia = False,
             has_rocm = False,
             rocm_gfx_target = None,
+            has_intel_gpu = False,
         )
     gfx = _normalize_forwarded_gfx(override_rocm_gfx)
     if gfx:
@@ -3135,21 +3133,6 @@ def _apply_host_overrides(
     return host
 
 
-def published_repo_for_host(host: HostInfo, *, linux_amd_tooling_present: bool = False) -> str:
-    """The release repo setup.sh / setup.ps1 pick for this host: macOS always the
-    fork (ggml-org macOS bundles need too-new macOS); else CPU-only Linux/Windows
-    -> ggml-org upstream (the fork ships no CPU bundle) and any usable GPU (NVIDIA
-    or ROCm) -> the fork. linux_amd_tooling_present mirrors setup.sh routing Linux
-    hosts that expose AMD tooling (rocminfo/amd-smi/hipconfig/hipinfo) to the fork
-    even when the probe cannot confirm an active GPU. Mirrors the shell routing."""
-    if host.is_macos:
-        return DEFAULT_PUBLISHED_REPO
-    has_gpu = (
-        host.has_usable_nvidia or host.has_rocm or (host.is_linux and linux_amd_tooling_present)
-    )
-    return DEFAULT_PUBLISHED_REPO if has_gpu else UPSTREAM_REPO
-
-
 def pick_windows_cuda_runtime(host: HostInfo) -> str | None:
     if not host.driver_cuda_version:
         return None
@@ -3881,6 +3864,23 @@ def resolve_upstream_asset_choice(host: HostInfo, llama_tag: str) -> AssetChoice
                 "falling back to source build with HIP support"
             )
 
+        # Intel (or other non-NVIDIA/non-AMD) GPU: use the Vulkan prebuilt. No
+        # physical NVIDIA (not just no usable one): a CUDA-hidden card must not
+        # be reached through Vulkan, which ignores CUDA_VISIBLE_DEVICES.
+        if host.has_intel_gpu and not host.has_physical_nvidia and not host.has_rocm:
+            vulkan_name = f"llama-{llama_tag}-bin-ubuntu-vulkan-x64.tar.gz"
+            if vulkan_name in upstream_assets:
+                log(f"Intel GPU detected -- using upstream Vulkan prebuilt {vulkan_name}")
+                return AssetChoice(
+                    repo = UPSTREAM_REPO,
+                    tag = llama_tag,
+                    name = vulkan_name,
+                    url = upstream_assets[vulkan_name],
+                    source_label = "upstream",
+                    install_kind = "linux-vulkan",
+                )
+            log("Intel GPU detected but no Vulkan prebuilt found -- falling back to CPU")
+
         upstream_name = f"llama-{llama_tag}-bin-ubuntu-x64.tar.gz"
         if upstream_name not in upstream_assets:
             raise PrebuiltFallback("upstream Linux CPU asset was not found")
@@ -3923,6 +3923,24 @@ def resolve_upstream_asset_choice(host: HostInfo, llama_tag: str) -> AssetChoice
                 )
             log("AMD ROCm detected on Windows but no HIP prebuilt found -- falling back to CPU")
 
+        # Intel (or other non-NVIDIA/non-AMD) GPU on Windows: use Vulkan. No
+        # physical NVIDIA so a CUDA-hidden card isn't reached through Vulkan.
+        if host.has_intel_gpu and not host.has_physical_nvidia and not host.has_rocm:
+            vulkan_name = f"llama-{llama_tag}-bin-win-vulkan-x64.zip"
+            if vulkan_name in upstream_assets:
+                log(
+                    f"Intel GPU detected on Windows -- using upstream Vulkan prebuilt {vulkan_name}"
+                )
+                return AssetChoice(
+                    repo = UPSTREAM_REPO,
+                    tag = llama_tag,
+                    name = vulkan_name,
+                    url = upstream_assets[vulkan_name],
+                    source_label = "upstream",
+                    install_kind = "windows-vulkan",
+                )
+            log("Intel GPU detected on Windows but no Vulkan prebuilt found -- falling back to CPU")
+
         upstream_name = f"llama-{llama_tag}-bin-win-cpu-x64.zip"
         if upstream_name not in upstream_assets:
             raise PrebuiltFallback("upstream Windows CPU asset was not found")
@@ -4015,6 +4033,9 @@ def resolve_release_asset_choice(
             published_choice = published_rocm_choice_for_host(release, host, "windows-rocm")
         else:
             published_choice = published_asset_choice_for_kind(release, "windows-cpu")
+    elif host.is_windows and host.is_arm64:
+        # Windows arm64 has no GPU prebuilt, so it always takes the CPU bundle.
+        published_choice = published_asset_choice_for_kind(release, "windows-arm64")
     elif host.is_macos and host.is_arm64:
         published_choice = published_asset_choice_for_kind(release, "macos-arm64")
     elif host.is_macos and host.is_x86_64:
@@ -4515,6 +4536,7 @@ def runtime_patterns_for_choice(choice: AssetChoice) -> list[str]:
         "linux-arm64-cuda",
         "linux-rocm",
         "linux-arm64",
+        "linux-vulkan",
     }:
         return ["llama-server", "llama-quantize", "llama-diffusion-gemma-visual-server", "lib*.so*"]
     if choice.install_kind in {"macos-arm64", "macos-x64"}:
@@ -4528,6 +4550,7 @@ def runtime_patterns_for_choice(choice: AssetChoice) -> list[str]:
         "windows-cpu",
         "windows-cuda",
         "windows-hip",
+        "windows-vulkan",
         "windows-rocm",
         "windows-arm64",
     }:
@@ -5743,8 +5766,10 @@ def validate_server(
             "linux-cuda",
             "linux-arm64-cuda",
             "linux-rocm",
+            "linux-vulkan",
             "windows-cuda",
             "windows-hip",
+            "windows-vulkan",
             "windows-rocm",
             "macos-arm64",
         }
@@ -6127,8 +6152,13 @@ def _linux_published_attempts(host: HostInfo, bundle: PublishedReleaseBundle) ->
         # CPU-only host. A usable-NVIDIA host never reaches here -- if its CUDA
         # selection produced nothing we want an empty attempt list so the caller
         # source-builds with CUDA, not a CPU-only binary silently installed on a
-        # GPU host (mirrors the ROCm branch, and Windows NVIDIA).
-        cpu_choice = published_asset_choice_for_kind(bundle, "linux-cpu")
+        # GPU host (mirrors the ROCm branch, and Windows NVIDIA). Only x86_64 and
+        # arm64 have a CPU bundle; any other Linux arch (ppc64le, riscv64, s390x)
+        # has none, so leave attempts empty and source-build rather than hand it
+        # the x86_64 linux-cpu binary (the Linux preflight checks libraries, not
+        # ELF arch, so a wrong-arch binary would not be caught).
+        kind = "linux-cpu" if host.is_x86_64 else "linux-arm64" if host.is_arm64 else None
+        cpu_choice = published_asset_choice_for_kind(bundle, kind) if kind else None
         if cpu_choice is not None:
             attempts.append(cpu_choice)
     return attempts
@@ -6143,9 +6173,9 @@ def _fork_manifest_release_plans(
     max_release_fallbacks: int = DEFAULT_MAX_PREBUILT_RELEASE_FALLBACKS,
 ) -> tuple[str, list[InstallReleasePlan]]:
     """Manifest-reading branch of resolve_simple_install_release_plans, used for
-    the fork's bundles whose GPU/arch coverage lives in
-    llama-prebuilt-manifest.json rather than in the filename: arm64 CUDA, Windows
-    CUDA, per-gfx ROCm, and macOS. Linux x64 takes the faster filename path."""
+    every fork host: all of the fork's bundles describe their GPU/arch coverage
+    in llama-prebuilt-manifest.json rather than in the asset filename (CPU,
+    x64/arm64 CUDA, Windows CUDA, per-gfx ROCm, and macOS)."""
     requested_tag = normalized_requested_llama_tag(llama_tag)
     allow_older_release_fallback = requested_tag == "latest" and not published_release_tag
     release_limit = max(1, max_release_fallbacks)
@@ -6361,6 +6391,20 @@ def runtime_payload_health_groups(choice: AssetChoice) -> list[list[str]]:
             ["libmtmd.so*"],
             ["libggml-hip.so*"],
         ]
+    if choice.install_kind == "linux-vulkan":
+        return [
+            ["libllama-common.so*"],
+            ["libllama.so*"],
+            ["libggml.so*"],
+            ["libggml-base.so*"],
+            # Match the sibling globs (linux-cuda/-rocm): x64 bundles ship
+            # arch-suffixed libggml-cpu-<variant>.so, arm64 may ship a bare
+            # libggml-cpu.so; the '-' form missed the latter and re-flagged
+            # the install unhealthy on every check.
+            ["libggml-cpu*.so*"],
+            ["libmtmd.so*"],
+            ["libggml-vulkan.so*"],
+        ]
     if choice.install_kind in {"windows-cpu", "windows-arm64"}:
         return [["llama.dll"]]
     if choice.install_kind == "windows-cuda":
@@ -6380,6 +6424,8 @@ def runtime_payload_health_groups(choice: AssetChoice) -> list[list[str]]:
         return groups
     if choice.install_kind in {"windows-hip", "windows-rocm"}:
         return [["llama.dll"], ["*hip*.dll"]]
+    if choice.install_kind == "windows-vulkan":
+        return [["llama.dll"], ["ggml-vulkan.dll"]]
     return []
 
 
@@ -6661,6 +6707,89 @@ def validate_prebuilt_attempts(
     raise PrebuiltFallback("no prebuilt bundle passed validation")
 
 
+def force_vulkan_requested() -> bool:
+    """Whether UNSLOTH_FORCE_VULKAN opts this host into the Vulkan llama.cpp
+    prebuilt instead of its detected CUDA/ROCm backend (e.g. so an AMD user can
+    run the Vulkan build for inference). Scoped to the llama.cpp backend; the
+    torch/training stack installs separately and still sees the real GPU.
+    """
+    return os.environ.get("UNSLOTH_FORCE_VULKAN", "").strip().lower() in (
+        "1",
+        "true",
+        "yes",
+    )
+
+
+def _vulkan_only_host(host: HostInfo) -> HostInfo:
+    """Rewrite ``host`` so the asset selectors take their Vulkan branch.
+
+    That branch fires on ``has_intel_gpu and not nvidia and not rocm``, so clear
+    the CUDA/ROCm flags and raise the integrated-GPU flag. The synthetic flag
+    never leaves install planning -- it only routes the llama.cpp prebuilt
+    choice, not the torch/training stack.
+    """
+    return dataclasses_replace(
+        host,
+        has_usable_nvidia = False,
+        has_physical_nvidia = False,
+        has_rocm = False,
+        has_intel_gpu = True,
+    )
+
+
+def _route_to_vulkan_prebuilt(
+    host: HostInfo, published_repo: str, published_release_tag: str, *, force_cpu: bool
+) -> tuple[HostInfo, str, str]:
+    """Point a Vulkan-capable host at the upstream ggml-org Vulkan prebuilt.
+
+    The unsloth published repo ships only CUDA/ROCm/CPU assets, so Vulkan comes
+    from UPSTREAM_REPO. Two triggers route here, both suppressed under
+    --cpu-fallback (the explicit "give me CPU" last resort wins):
+      * UNSLOTH_FORCE_VULKAN forces Vulkan over the detected CUDA/ROCm backend;
+      * an auto-detected Intel GPU with NO physical NVIDIA/ROCm -- the purpose
+        of the has_intel_gpu probe, since the fork manifest ships no Vulkan asset.
+    Applied by BOTH the install path and the --resolve-prebuilt probe so the
+    "is a prebuilt available" answer matches what actually gets installed.
+
+    Returns the (possibly rewritten) host, repo, and release tag.
+    """
+    forced = force_vulkan_requested()
+    # Gate auto-routing on no PHYSICAL NVIDIA, not merely no usable one: a mixed
+    # NVIDIA+Intel host that hides NVIDIA with CUDA_VISIBLE_DEVICES=""/-1 keeps
+    # has_physical_nvidia=True while has_usable_nvidia goes False. Vulkan ignores
+    # CUDA_VISIBLE_DEVICES, so auto-routing such a host would let it grab the
+    # reserved NVIDIA GPU. An explicit UNSLOTH_FORCE_VULKAN still overrides.
+    auto_intel = host.has_intel_gpu and not host.has_physical_nvidia and not host.has_rocm
+    if force_cpu or not (forced or auto_intel):
+        return host, published_repo, published_release_tag
+    if host.is_macos:
+        if forced:
+            log(
+                "UNSLOTH_FORCE_VULKAN is set but ignored on macOS "
+                "(Metal is used; there is no Vulkan prebuilt)"
+            )
+        return host, published_repo, published_release_tag
+    if forced:
+        log(
+            "UNSLOTH_FORCE_VULKAN is set; installing the upstream Vulkan "
+            "llama.cpp prebuilt instead of the detected GPU backend"
+        )
+        # Forcing may override a detected NVIDIA/ROCm host, so normalize it to
+        # Vulkan-only; an auto-detected Intel host already is.
+        host = _vulkan_only_host(host)
+    else:
+        log("Intel GPU detected; installing the upstream Vulkan llama.cpp prebuilt")
+    # Swapping the fork for upstream invalidates a fork release pin: the two use
+    # different tag namespaces (fork b9596-mix-<sha> vs upstream b9596), so a
+    # pinned fork tag would make the upstream resolver query a nonexistent
+    # release and fall back to source. Drop it and let the upstream resolver
+    # pick by the requested llama tag. A pin already on an explicit upstream repo
+    # (repo unchanged here) is preserved.
+    if published_repo != UPSTREAM_REPO:
+        published_release_tag = ""
+    return host, UPSTREAM_REPO, published_release_tag
+
+
 def diffusion_visual_server_backfill_needed(
     install_dir: Path, host: HostInfo, choice: AssetChoice
 ) -> bool:
@@ -6703,6 +6832,9 @@ def install_prebuilt(
         override_rocm_gfx = override_rocm_gfx,
         force_cpu = force_cpu,
     )
+    host, published_repo, published_release_tag = _route_to_vulkan_prebuilt(
+        host, published_repo, published_release_tag, force_cpu = force_cpu
+    )
     choice: AssetChoice | None = None
     try:
         with install_lock(install_lock_path(install_dir)):
@@ -6714,8 +6846,10 @@ def install_prebuilt(
                 log(
                     f"no existing llama.cpp install detected at {install_dir}; performing fresh prebuilt install"
                 )
-            # Single resolver: linux-x64 takes the fast filename path internally,
-            # every other fork host reads the manifest.
+            # Single resolver: every fork host selects from the release manifest;
+            # an explicit ggml-org override selects by asset filename instead. A
+            # forced-Vulkan host already has published_repo pointed at
+            # UPSTREAM_REPO above, so the resolver takes the Vulkan asset branch.
             requested_tag, release_plans = resolve_simple_install_release_plans(
                 llama_tag,
                 host,
@@ -6903,8 +7037,8 @@ def parse_args() -> argparse.Namespace:
         const = "latest",
         help = (
             "Report whether an official prebuilt exists for this host without "
-            "downloading. Picks the host's published repo when --published-repo "
-            "is left at the default. Use --output-format json."
+            "downloading. Plans against --published-repo (defaults to the "
+            "fork). Use --output-format json."
         ),
     )
     parser.add_argument(
@@ -6992,27 +7126,23 @@ def main() -> int:
         return EXIT_SUCCESS
 
     if args.resolve_prebuilt is not None:
-        # Host-aware "is a prebuilt available" probe, no download. A default repo
-        # means "pick the repo for this host"; PrebuiltFallback == source build.
+        # Host-aware "is a prebuilt available" probe, no download. Every host now
+        # plans against the fork (args.published_repo defaults to it); an explicit
+        # --published-repo overrides. PrebuiltFallback == source build.
         host = _apply_host_overrides(
             detect_host(),
             override_has_rocm = args.has_rocm,
             override_rocm_gfx = args.rocm_gfx,
             force_cpu = args.cpu_fallback,
         )
-        # setup.sh routes Linux hosts with AMD tooling to the fork even when no GPU
-        # is probed; mirror that so a HIP source build is not offered a CPU prebuilt.
-        amd_tooling = host.is_linux and any(
-            shutil.which(t) for t in ("rocminfo", "amd-smi", "hipconfig", "hipinfo")
-        )
-        repo = (
-            published_repo_for_host(host, linux_amd_tooling_present = amd_tooling)
-            if args.published_repo == DEFAULT_PUBLISHED_REPO
-            else args.published_repo
+        # Same Vulkan routing the install path applies, so the probe's answer
+        # matches what would install (an Intel/forced-Vulkan host -> upstream).
+        host, repo, release_tag = _route_to_vulkan_prebuilt(
+            host, args.published_repo, args.published_release_tag or "", force_cpu = args.cpu_fallback
         )
         try:
             _requested, plans = resolve_simple_install_release_plans(
-                args.resolve_prebuilt, host, repo, args.published_release_tag or ""
+                args.resolve_prebuilt, host, repo, release_tag
             )
             choice = plans[0].attempts[0] if plans and plans[0].attempts else None
             if choice is None:
diff --git a/studio/install_python_stack.py b/studio/install_python_stack.py
index e033a56a0a..19c492deaa 100644
--- a/studio/install_python_stack.py
+++ b/studio/install_python_stack.py
@@ -103,33 +103,47 @@ _PYTORCH_WHL_BASE = (
     os.environ.get("UNSLOTH_PYTORCH_MIRROR") or "https://download.pytorch.org/whl"
 ).rstrip("/")
 
-# CUDA torch repair specs (see _ensure_cuda_torch). torchvision/torchaudio are
-# pinned to the torch<2.11 family rather than left bare: the install uses an
-# exclusive --index-url (no PyPI fallback), so a bare name could resolve a
-# torchvision built against a different torch major (e.g. 0.27 for torch 2.12)
-# and fail at runtime with an ABI mismatch. Same bounds as the _default ROCm
-# spec above, which targets the same torch family.
+# CUDA torch repair specs (see _ensure_cuda_torch). torch 2.11 is allowed: its
+# torchao 0.17 cpp kernels load cleanly (0.16 crashes on cu130), and the flash-attn
+# / causal-conv1d / mamba torch2.10 wheels load and pass their upstream suites on
+# 2.11 (see wheel_utils._PREBUILT_WHEEL_TORCH_MM). torchvision/torchaudio are pinned
+# (not bare) because the install uses an exclusive --index-url (no PyPI fallback), so
+# a bare name could resolve one built against a different torch major (e.g. 0.27 for
+# torch 2.12) and fail at runtime with an ABI mismatch.
 _CUDA_TORCH_PKG_SPEC: tuple[str, str, str] = (
-    "torch>=2.4,<2.11.0",
-    "torchvision>=0.19,<0.26.0",
-    "torchaudio>=2.4,<2.11.0",
+    "torch>=2.4,<2.12.0",
+    "torchvision>=0.19,<0.27.0",
+    "torchaudio>=2.4,<2.12.0",
 )
 
-# torchao's C++ extensions are built against ONE exact torch release; a newer
-# torch makes torchao skip its cpp kernels ("Skipping import of cpp extensions
-# due to incompatible torch version ...") and fall back to slow Python. Because
-# the torch pin above is a range (and every CUDA index now tops out at torch
-# 2.10), the torch actually installed drifts ahead of a fixed torchao pin. So
-# pick the torchao whose build matches the torch in the venv. Table: pytorch/ao#2919.
-#   torch 2.9.x  -> torchao 0.14.0 (today's pin; built for torch 2.9.0)
-#   torch 2.10.x -> torchao 0.16.0 (built for torch 2.10.0)
-#   torch 2.11.x -> torchao 0.17.0 (built for torch 2.11.0; reachable via ROCm rocm7.2)
-# Unknown/older torch keeps the conservative default (no regression vs today).
+# torchao's cpp extensions are pinned to ONE torch release AND CUDA major. A torch
+# mismatch just skips the cpp kernels (slow Python fallback); a CUDA mismatch fails
+# to import ("libcudart.so.12: cannot open shared object file"). The torch pin is a
+# range, so match torchao to the installed torch (table: pytorch/ao#2919):
+#   2.9.x            -> 0.14.0
+#   2.10.x, CUDA<=12 -> 0.16.0 (cpp built for 2.10, loads via the CUDA-12 wheel)
+#   2.10.x, CUDA>=13 -> 0.17.0 (cu130: 0.16.0's CUDA-12 cpp crashes on load; 0.17.0
+#                       targets torch 2.11 so its cpp is cleanly skipped, not crashed)
+#   2.11.x           -> 0.17.0 (reachable via CUDA or ROCm rocm7.2)
+# Unknown/older torch keeps the conservative default.
 _TORCHAO_DEFAULT_SPEC = "torchao==0.14.0"
-_TORCHAO_BY_TORCH_MINOR: dict[int, str] = {
-    10: "torchao==0.16.0",
-    11: "torchao==0.17.0",
-}
+_TORCHAO_TORCH_210_SPEC = "torchao==0.16.0"
+_TORCHAO_TORCH_210_CUDA13_SPEC = "torchao==0.17.0"
+_TORCHAO_TORCH_211_PLUS_SPEC = "torchao==0.17.0"
+# torch 2.10 built against CUDA >= this major can't load 0.16.0's CUDA-12 cpp.
+_TORCHAO_CUDA13_MIN_MAJOR = 13
+
+
+def _cuda_major_from_torch_version(torch_version: str) -> int | None:
+    """Extract the CUDA major from a torch local version tag, e.g. '2.10.0+cu130'
+    -> 13, '2.10.0+cu128' -> 12. Returns None for rocm/cpu/tagless builds."""
+    local = str(torch_version).split("+", 1)
+    if len(local) < 2 or not local[1].startswith("cu"):
+        return None
+    digits = re.sub(r"[^0-9].*", "", local[1][2:])  # 'cu130' -> '130'
+    if not digits:
+        return None
+    return int(digits) // 10  # '130' -> 13, '128' -> 12, '118' -> 11
 
 
 def _select_torchao_spec(torch_version: str | None) -> str:
@@ -151,8 +165,14 @@ def _select_torchao_spec(torch_version: str | None) -> str:
     if major != 2:
         return _TORCHAO_DEFAULT_SPEC
     if minor >= 11:
-        return _TORCHAO_BY_TORCH_MINOR[11]  # newest known build; covers 2.11+
-    return _TORCHAO_BY_TORCH_MINOR.get(minor, _TORCHAO_DEFAULT_SPEC)
+        return _TORCHAO_TORCH_211_PLUS_SPEC  # newest known build; covers 2.11+
+    if minor == 10:
+        # cu130+ can't load 0.16.0's CUDA-12 cpp; use 0.17.0 (cpp skipped, not crashed).
+        cuda_major = _cuda_major_from_torch_version(str(torch_version))
+        if cuda_major is not None and cuda_major >= _TORCHAO_CUDA13_MIN_MAJOR:
+            return _TORCHAO_TORCH_210_CUDA13_SPEC
+        return _TORCHAO_TORCH_210_SPEC
+    return _TORCHAO_DEFAULT_SPEC
 
 
 def _probe_installed_torch_version() -> str | None:
diff --git a/studio/setup.ps1 b/studio/setup.ps1
index bb1e88cc4e..db01a1ecad 100644
--- a/studio/setup.ps1
+++ b/studio/setup.ps1
@@ -2621,7 +2621,18 @@ function Fast-Install {
     param([Parameter(ValueFromRemainingArguments=$true)]$Args_)
     if ($UseUv) {
         $VenvPy = (Get-Command python).Source
-        $result = & uv pip install --python $VenvPy @Args_ 2>&1
+        # An explicit --index-url must win. Inherited uv index env vars otherwise
+        # override it and pull CPU torch over the CUDA/ROCm build (#6898), so drop
+        # them only for index-pinned installs; mirrors still apply elsewhere.
+        $saved = @{}
+        if (@($Args_) -contains '--index-url') {
+            foreach ($n in 'UV_DEFAULT_INDEX', 'UV_INDEX_URL', 'UV_INDEX', 'UV_EXTRA_INDEX_URL') {
+                $saved[$n] = [Environment]::GetEnvironmentVariable($n)
+                Remove-Item "Env:$n" -ErrorAction SilentlyContinue
+            }
+        }
+        try { $result = & uv pip install --python $VenvPy @Args_ 2>&1 }
+        finally { foreach ($n in $saved.Keys) { if ($null -ne $saved[$n]) { Set-Item "Env:$n" $saved[$n] } } }
         if ($LASTEXITCODE -eq 0) { return }
     }
     & python -m pip install @Args_ 2>&1
@@ -3088,12 +3099,11 @@ $LlamaCppDir = Join-Path $UnslothHome "llama.cpp"
 $NeedLlamaSourceBuild = $false
 $SkipPrebuiltInstall = $false
 $RequestedLlamaTag = if ($env:UNSLOTH_LLAMA_TAG) { $env:UNSLOTH_LLAMA_TAG } else { $DefaultLlamaTag }
-# GPU Windows (CUDA / ROCm) installs the fork's app-* prebuilts; CPU-only stays
-# on ggml-org (the fork ships no windows-cpu bundle). Mirrors setup.sh's routing.
-# A resolved gfx arch counts as a GPU host even when $HasROCm is false (Adrenalin
-# driver only, no HIP runtime): the fork's per-gfx bundle ships its own runtime,
-# so route there instead of ggml-org / a CPU build.
-$HelperReleaseRepo = if ($HasNvidiaSmi -or $HasROCm -or $script:ROCmGfxArch) { "unslothai/llama.cpp" } else { "ggml-org/llama.cpp" }
+# Every host installs the fork's app-* prebuilts now: GPU Windows (CUDA / ROCm)
+# already did, and the fork now also ships the CPU bundles for Windows x64 and
+# arm64 (windows-cpu / windows-arm64). ggml-org artifacts are no longer used by
+# default. Mirrors setup.sh's routing.
+$HelperReleaseRepo = "unslothai/llama.cpp"
 $LlamaPr = if ($env:UNSLOTH_LLAMA_PR) { $env:UNSLOTH_LLAMA_PR.Trim() } else { "" }
 
 $LlamaPrForce = if ($env:UNSLOTH_LLAMA_PR_FORCE) { $env:UNSLOTH_LLAMA_PR_FORCE.Trim() } else { $DefaultLlamaPrForce }
@@ -3283,11 +3293,13 @@ if ($LocalLlamaCppLinked) {
                 # treat a valid ROCm install as mismatched. A name-inferred gfx
                 # arch (Adrenalin-only, no confirmed runtime) still counts as
                 # ROCm-capable -- the ROCm prebuilt bundles its own runtime,
-                # mirroring the --rocm-gfx forward below. NOTE: this block is
-                # currently inert -- write_prebuilt_metadata does not persist an
-                # install_kind key, so $existingKind is always null. If that changes,
-                # add the remaining host kinds (e.g. windows-arm64) before relying on it.
-                $expectedKinds = if ($HasROCm -or $script:ROCmGfxArch) { @("windows-rocm", "windows-hip") } elseif ($HasNvidiaSmi) { @("windows-cuda") } else { @("windows-cpu") }
+                # mirroring the --rocm-gfx forward below. The CPU branch covers both
+                # the x64 windows-cpu and arm64 windows-arm64 bundles (Windows arm64
+                # has no GPU prebuilt). NOTE: this block is currently inert --
+                # write_prebuilt_metadata does not persist an install_kind key, so
+                # $existingKind is always null; keep $expectedKinds in sync with the
+                # kinds install_llama_prebuilt.py installs before relying on it.
+                $expectedKinds = if ($HasROCm -or $script:ROCmGfxArch) { @("windows-rocm", "windows-hip") } elseif ($HasNvidiaSmi) { @("windows-cuda") } else { @("windows-cpu", "windows-arm64") }
                 if ($existingKind -and ($existingKind -notin $expectedKinds)) {
                     substep "Removing mismatched llama.cpp install (found '$existingKind', need one of: $($expectedKinds -join ', '))..."
                     Remove-Item -Recurse -Force -LiteralPath $LlamaCppDir -ErrorAction SilentlyContinue
diff --git a/studio/setup.sh b/studio/setup.sh
index 6a74cd2296..d244e3cdcf 100755
--- a/studio/setup.sh
+++ b/studio/setup.sh
@@ -1176,63 +1176,19 @@ _REQUESTED_LLAMA_TAG="${UNSLOTH_LLAMA_TAG:-${_DEFAULT_LLAMA_TAG}}"
 _HOST_SYSTEM="$(uname -s 2>/dev/null || true)"
 _HOST_MACHINE="$(uname -m 2>/dev/null || true)"
 
-# Pick the release repo install_llama_prebuilt.py plans against.
-# The fork ships CUDA (Linux x64/arm64, Windows), ROCm (Linux/Windows) and
-# macOS bundles. Only the plain CPU/Vulkan bundles still come from ggml-org, so
-# CPU-only Linux (x86_64 and arm64) routes there; GPU Linux, Windows and macOS
-# use unslothai.
-_LINUX_HAS_GPU=false
-# Route to the fork only for a usable GPU. NVIDIA counts only when a device is
-# actually enumerated and not hidden via CUDA_VISIBLE_DEVICES=""/-1
-# (_setup_nvidia_usable, from _setup_has_usable_nvidia_gpu above) -- mirroring
-# install_llama_prebuilt.py's has_usable_nvidia. Mere nvidia-smi presence
-# (CPU-only CUDA-toolkit containers, broken drivers) or a hidden GPU therefore
-# takes the ggml-org CPU prebuilt instead of a slow source build. AMD is
-# deliberately left on tooling presence, not usability: an unusable NVIDIA host
-# has a good CPU prebuilt to fall back to, whereas tightening AMD would regress
-# ROCm hosts exposing only hipconfig/hipinfo into an unnecessary CPU build.
-if [ "$_setup_nvidia_usable" = true ]; then
-    _LINUX_HAS_GPU=true
-else
-    for _GPU_TOOL in rocminfo amd-smi hipconfig hipinfo; do
-        if command -v "$_GPU_TOOL" >/dev/null 2>&1; then
-            _LINUX_HAS_GPU=true
-            break
-        fi
-    done
-fi
+# Pick the release repo install_llama_prebuilt.py plans against. Every host this
+# installer supports now pulls its llama.cpp prebuilt from the unslothai fork: it
+# ships the CUDA (Linux x64/arm64, Windows), ROCm (Linux/Windows) and macOS
+# bundles, plus the CPU bundles for Linux/Windows on both x86_64 and arm64.
+# ggml-org artifacts are no longer used by default.
+_HELPER_RELEASE_REPO="unslothai/llama.cpp"
 # UNSLOTH_ROCM_GFX_ARCH may be set on a host where no probe fired, so the override
 # nested in the AMD-detected branch above never ran and _setup_gfx is still empty.
-# Honour it here so the routing guard below and the --rocm-gfx forwarding both see
-# it (install_llama_prebuilt.py reads the same env var as the --rocm-gfx default).
-if [ "$_setup_nvidia_usable" != true ] && [ -z "${_setup_gfx:-}" ] && [ -n "${UNSLOTH_ROCM_GFX_ARCH:-}" ]; then
+# Honour it here so the --rocm-gfx forwarding below still sees it
+# (install_llama_prebuilt.py reads the same env var as the --rocm-gfx default).
+if [ "${_setup_nvidia_usable:-}" != true ] && [ -z "${_setup_gfx:-}" ] && [ -n "${UNSLOTH_ROCM_GFX_ARCH:-}" ]; then
     _setup_gfx="${UNSLOTH_ROCM_GFX_ARCH}"
 fi
-# A resolved/forwarded gfx arch (UNSLOTH_ROCM_GFX_ARCH) means an AMD GPU even when
-# no ROCm tooling is on PATH; route it to the fork so the per-gfx prebuilt is
-# picked instead of ggml-org / a source build.
-if [ "$_LINUX_HAS_GPU" = false ] && [ -n "${_setup_gfx:-}" ]; then
-    _LINUX_HAS_GPU=true
-fi
-
-if [ "$_HOST_SYSTEM" = "Linux" ] \
-        && [ "$_HOST_MACHINE" = "x86_64" ] \
-        && [ "$_LINUX_HAS_GPU" = false ]; then
-    _HELPER_RELEASE_REPO="ggml-org/llama.cpp"
-elif [ "$_HOST_SYSTEM" = "Linux" ] \
-        && { [ "$_HOST_MACHINE" = "aarch64" ] || [ "$_HOST_MACHINE" = "arm64" ]; } \
-        && [ "$_LINUX_HAS_GPU" = false ]; then
-    # CPU-only Linux ARM64 (Ampere Altra, Raspberry Pi 5, GitHub
-    # `ubuntu-24.04-arm`, CPU-only Jetson rescue mode, ...). The fork ships no
-    # arm64 CPU bundle, so without this branch the prebuilt resolver returns 0
-    # attempts and the installer falls back to a source build. ggml-org ships
-    # llama-bNNNN-bin-ubuntu-arm64.tar.gz from at least b9072 onward.
-    _HELPER_RELEASE_REPO="ggml-org/llama.cpp"
-else
-    # GPU Linux (x64 CUDA/ROCm, arm64 CUDA), Windows (CUDA/ROCm), and macOS.
-    _HELPER_RELEASE_REPO="unslothai/llama.cpp"
-fi
-unset _GPU_TOOL
 _LLAMA_PR="${UNSLOTH_LLAMA_PR:-}"
 _SKIP_PREBUILT_INSTALL=false
 _LLAMA_PR_FORCE="${UNSLOTH_LLAMA_PR_FORCE:-${_DEFAULT_LLAMA_PR_FORCE}}"
@@ -1939,19 +1895,19 @@ else
 fi  # end _SKIP_GGUF_BUILD check
 
 # ── arm64 Linux GPU: CPU prebuilt as a last resort ──
-# arm64 Linux with a GPU has no CUDA prebuilt anywhere (the unslothai fork is
-# x64 only; ggml-org ships no Linux CUDA build), so it source-builds for the
-# GPU above. If that produced no binary, install ggml-org's arm64 CPU prebuilt
-# instead of leaving the host without llama.cpp.
+# An arm64 Linux GPU host source-builds for the GPU above. If that produced no
+# binary, install the fork's arm64 CPU prebuilt (app-<tag>-linux-arm64-cpu.tar.gz)
+# instead of leaving the host without llama.cpp. --cpu-fallback drops the GPU
+# attributes so the CPU bundle is selected rather than re-attempting CUDA.
 if [ "$_LLAMA_CPP_DEGRADED" = true ] \
         && [ "$_HOST_SYSTEM" = "Linux" ] \
         && { [ "$_HOST_MACHINE" = "aarch64" ] || [ "$_HOST_MACHINE" = "arm64" ]; }; then
-    substep "GPU source build unavailable; trying ggml-org arm64 CPU prebuilt..."
+    substep "GPU source build unavailable; trying arm64 CPU prebuilt..."
     _ARM64_CPU_CMD=(
         python "$SCRIPT_DIR/install_llama_prebuilt.py"
         --install-dir "$LLAMA_CPP_DIR"
         --llama-tag "$_REQUESTED_LLAMA_TAG"
-        --published-repo "ggml-org/llama.cpp"
+        --published-repo "unslothai/llama.cpp"
         --cpu-fallback
     )
     # Trust the installer's exit code: it validates the server before exiting 0,
diff --git a/tests/python/test_fast_sentence_transformer_embedding_parity.py b/tests/python/test_fast_sentence_transformer_embedding_parity.py
new file mode 100644
index 0000000000..252d4486a5
--- /dev/null
+++ b/tests/python/test_fast_sentence_transformer_embedding_parity.py
@@ -0,0 +1,122 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team.
+"""Regression guard for issue #6881: FastSentenceTransformer must preprocess text
+like a stock SentenceTransformer for decoder embedding models. ST 5.x infers a
+"message" modality for chat-template models (e.g. Qwen/Qwen3-Embedding), so building
+via `Transformer(model_name, ...)` chat-wraps inputs and degrades embeddings;
+`_create_transformer_module` uses `Transformer.load(...)` instead.
+
+Layers: test_transformer_load_signature_supports_unsloth_kwargs (fast, runs when ST
+is importable) and test_fast_sentence_transformer_matches_stock_st (end-to-end parity,
+opt-in via UNSLOTH_EMBEDDING_PARITY_MODEL so default CI is unaffected).
+"""
+
+from __future__ import annotations
+
+import inspect
+import os
+
+import pytest
+
+
+def test_transformer_load_signature_supports_unsloth_kwargs():
+    """Forwards-compat tripwire: a Hub-capable Transformer.load must accept the kwargs
+    the #6881 fix passes. Legacy ST 3.x/4.x expose load(input_path); the code falls back
+    to Transformer(...) there, so mirror that gate and skip."""
+    models = pytest.importorskip("sentence_transformers.models")
+    load = getattr(models.Transformer, "load", None)
+    assert callable(load), (
+        "sentence_transformers Transformer.load is missing; the #6881 fix in "
+        "unsloth.models.sentence_transformer._create_transformer_module depends on it."
+    )
+    params = inspect.signature(load).parameters
+    accepts_var_kw = any(p.kind is inspect.Parameter.VAR_KEYWORD for p in params.values())
+    # Mirror _create_transformer_module's hub_capable gate.
+    hub_capable = accepts_var_kw or any(k in params for k in ("token", "cache_folder", "revision"))
+    if not hub_capable:
+        pytest.skip(
+            "legacy Transformer.load(input_path); production path falls back to Transformer(...)"
+        )
+    unsupported = [
+        k
+        for k in ("token", "cache_folder", "revision", "trust_remote_code")
+        if not (accepts_var_kw or k in params)
+    ]
+    assert not unsupported, (
+        f"installed sentence_transformers Transformer.load no longer accepts {unsupported} "
+        f"and has no **kwargs; update _create_transformer_module (#6881) before it silently "
+        f"falls back to Transformer(...)."
+    )
+
+
+def _probe_texts():
+    return [
+        "roasted chickpeas in 20 kg bags",
+        "The capital of France is Paris.",
+        "A fast brown fox jumps over the lazy dog.",
+        "recette de tarte aux pommes traditionnelle",
+    ]
+
+
+def test_fast_sentence_transformer_matches_stock_st():
+    """End-to-end: FastSentenceTransformer embeddings and tokenization must match a
+    stock SentenceTransformer load of the same checkpoint. Opt-in (needs a model) and
+    GPU-only (FastSentenceTransformer requires CUDA), so it skips on CPU-only runners."""
+    model_id = os.environ.get("UNSLOTH_EMBEDDING_PARITY_MODEL")
+    if not model_id:
+        pytest.skip(
+            "set UNSLOTH_EMBEDDING_PARITY_MODEL to a chat-template embedding model "
+            "(HF id or local path) to run the #6881 parity test"
+        )
+
+    torch = pytest.importorskip("torch")
+    if not torch.cuda.is_available():
+        pytest.skip("FastSentenceTransformer requires CUDA; skipping on CPU-only runner")
+    np = pytest.importorskip("numpy")
+    pytest.importorskip("sentence_transformers")
+    from sentence_transformers import SentenceTransformer
+
+    device = "cuda"
+    # Prefer bf16 when the GPU supports it: fp16 overflows to NaN on bf16-native
+    # embedders such as EmbeddingGemma (Gemma3), which would mask real parity.
+    dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
+    texts = _probe_texts()
+    max_seq_length = 256
+
+    # Control FIRST, before importing unsloth, so its global import patches never
+    # touch the stock reference (mirrors the issue's "restart runtime" repro).
+    ctrl = SentenceTransformer(model_id, device = device, model_kwargs = {"torch_dtype": dtype})
+    ctrl.max_seq_length = max_seq_length
+    ctrl_ids = ctrl.tokenize([texts[0]])["input_ids"][0].tolist()
+    ctrl_emb = np.asarray(
+        ctrl.encode(texts, normalize_embeddings = True, batch_size = 8), dtype = np.float32
+    )
+
+    import unsloth  # noqa: F401
+    from unsloth import FastSentenceTransformer
+
+    fast = FastSentenceTransformer.from_pretrained(
+        model_id,
+        max_seq_length = max_seq_length,
+        dtype = dtype,
+        load_in_4bit = False,
+        load_in_16bit = True,
+    )
+    fast_ids = fast.tokenize([texts[0]])["input_ids"][0].tolist()
+    fast_emb = np.asarray(
+        fast.encode(texts, normalize_embeddings = True, batch_size = 8), dtype = np.float32
+    )
+
+    # Identical tokenization = no chat-template wrapping slipped in (the #6881 defect).
+    assert fast_ids == ctrl_ids, (
+        f"tokenization diverged (chat-template wrapping regressed?):\n"
+        f"  stock: {ctrl_ids}\n  fast:  {fast_ids}"
+    )
+
+    cos = (ctrl_emb * fast_emb).sum(1) / (
+        np.linalg.norm(ctrl_emb, axis = 1) * np.linalg.norm(fast_emb, axis = 1)
+    )
+    assert float(cos.min()) > 0.99, (
+        f"embedding parity regressed: min cosine {float(cos.min()):.5f} <= 0.99 "
+        f"(per-text {[round(float(c), 5) for c in cos]})"
+    )
diff --git a/tests/python/test_flash_attn_install_python_stack.py b/tests/python/test_flash_attn_install_python_stack.py
index 26ff03505a..bf3ed57788 100644
--- a/tests/python/test_flash_attn_install_python_stack.py
+++ b/tests/python/test_flash_attn_install_python_stack.py
@@ -13,102 +13,35 @@ sys.path.insert(0, str(STUDIO_DIR))
 sys.path.insert(0, str(STUDIO_DIR / "backend"))
 
 import install_python_stack as ips
-from backend.utils import wheel_utils
+from utils import wheel_utils
 
 
-def _smi_result(stdout: str, returncode: int = 0) -> subprocess.CompletedProcess:
-    return subprocess.CompletedProcess(["nvidia-smi"], returncode, stdout, "")
+class TestPrebuiltWheelTorchMapping:
+    def test_torch_211_maps_to_torch210(self):
+        assert wheel_utils.prebuilt_wheel_torch_mm("2.11") == "2.10"
 
+    def test_other_versions_pass_through(self):
+        for torch_mm in ("2.9", "2.10", "2.12"):
+            assert wheel_utils.prebuilt_wheel_torch_mm(torch_mm) == torch_mm
 
-class TestHasBlackwellGpu:
-    def setup_method(self):
-        wheel_utils.has_blackwell_gpu.cache_clear()
-
-    def teardown_method(self):
-        wheel_utils.has_blackwell_gpu.cache_clear()
-
-    def test_returns_false_when_nvidia_smi_missing(self):
-        with mock.patch.object(wheel_utils.shutil, "which", return_value = None):
-            assert wheel_utils.has_blackwell_gpu() is False
-
-    def test_returns_true_for_sm_100(self):
-        with (
-            mock.patch.object(wheel_utils.shutil, "which", return_value = "/usr/bin/nvidia-smi"),
-            mock.patch.object(wheel_utils.subprocess, "run", return_value = _smi_result("10.0\n")),
-        ):
-            assert wheel_utils.has_blackwell_gpu() is True
-
-    def test_returns_true_for_sm_120(self):
-        with (
-            mock.patch.object(wheel_utils.shutil, "which", return_value = "/usr/bin/nvidia-smi"),
-            mock.patch.object(wheel_utils.subprocess, "run", return_value = _smi_result("12.0\n")),
-        ):
-            assert wheel_utils.has_blackwell_gpu() is True
-
-    def test_returns_true_for_sm_121(self):
-        with (
-            mock.patch.object(wheel_utils.shutil, "which", return_value = "/usr/bin/nvidia-smi"),
-            mock.patch.object(wheel_utils.subprocess, "run", return_value = _smi_result("12.1\n")),
-        ):
-            assert wheel_utils.has_blackwell_gpu() is True
-
-    def test_returns_false_for_sm_90(self):
-        with (
-            mock.patch.object(wheel_utils.shutil, "which", return_value = "/usr/bin/nvidia-smi"),
-            mock.patch.object(wheel_utils.subprocess, "run", return_value = _smi_result("9.0\n")),
-        ):
-            assert wheel_utils.has_blackwell_gpu() is False
-
-    def test_returns_false_for_sm_89(self):
-        with (
-            mock.patch.object(wheel_utils.shutil, "which", return_value = "/usr/bin/nvidia-smi"),
-            mock.patch.object(wheel_utils.subprocess, "run", return_value = _smi_result("8.9\n")),
-        ):
-            assert wheel_utils.has_blackwell_gpu() is False
-
-    def test_mixed_gpus_with_one_blackwell_returns_true(self):
-        with (
-            mock.patch.object(wheel_utils.shutil, "which", return_value = "/usr/bin/nvidia-smi"),
-            mock.patch.object(
-                wheel_utils.subprocess,
-                "run",
-                return_value = _smi_result("8.0\n10.0\n"),
-            ),
-        ):
-            assert wheel_utils.has_blackwell_gpu() is True
-
-    def test_returns_false_when_nvidia_smi_fails(self):
-        with (
-            mock.patch.object(wheel_utils.shutil, "which", return_value = "/usr/bin/nvidia-smi"),
-            mock.patch.object(
-                wheel_utils.subprocess,
-                "run",
-                return_value = _smi_result("", returncode = 1),
-            ),
-        ):
-            assert wheel_utils.has_blackwell_gpu() is False
-
-    def test_returns_false_on_subprocess_timeout(self):
-        with (
-            mock.patch.object(wheel_utils.shutil, "which", return_value = "/usr/bin/nvidia-smi"),
-            mock.patch.object(
-                wheel_utils.subprocess,
-                "run",
-                side_effect = subprocess.TimeoutExpired(cmd = "nvidia-smi", timeout = 10),
-            ),
-        ):
-            assert wheel_utils.has_blackwell_gpu() is False
-
-    def test_returns_false_on_malformed_output(self):
-        with (
-            mock.patch.object(wheel_utils.shutil, "which", return_value = "/usr/bin/nvidia-smi"),
-            mock.patch.object(
-                wheel_utils.subprocess,
-                "run",
-                return_value = _smi_result("not-a-number\n\n"),
-            ),
-        ):
-            assert wheel_utils.has_blackwell_gpu() is False
+    def test_direct_wheel_url_reuses_torch210_on_211(self):
+        # causal-conv1d / mamba go through direct_wheel_url; torch 2.11 reuses the
+        # torch2.10 wheel filename just like flash-attn does.
+        url = wheel_utils.direct_wheel_url(
+            filename_prefix = "causal_conv1d",
+            package_version = "1.6.1",
+            release_tag = "v1.6.1.post4",
+            release_base_url = "https://example.test/download",
+            env = {
+                "python_tag": "cp313",
+                "torch_mm": "2.11",
+                "cuda_major": "13",
+                "cxx11abi": "TRUE",
+                "platform_tag": "linux_x86_64",
+            },
+        )
+        assert url is not None
+        assert "causal_conv1d-1.6.1+cu13torch2.10cxx11abiTRUE-cp313-cp313-linux_x86_64.whl" in url
 
 
 class TestFlashAttnWheelSelection:
@@ -118,9 +51,24 @@ class TestFlashAttnWheelSelection:
     def test_torch_29_maps_to_v283(self):
         assert ips._select_flash_attn_version("2.9") == "2.8.3"
 
-    def test_unsupported_torch_has_no_wheel_mapping(self):
+    def test_torch_211_has_no_native_version_entry(self):
+        # The raw version table has no torch2.11-tagged wheel; the URL builder
+        # reuses the torch2.10 wheel instead (see test_torch_211_reuses_torch210_wheel).
         assert ips._select_flash_attn_version("2.11") is None
 
+    def test_torch_211_reuses_torch210_wheel(self):
+        url = ips._build_flash_attn_wheel_url(
+            {
+                "python_tag": "cp313",
+                "torch_mm": "2.11",
+                "cuda_major": "13",
+                "cxx11abi": "TRUE",
+                "platform_tag": "linux_x86_64",
+            }
+        )
+        assert url is not None
+        assert "flash_attn-2.8.1+cu13torch2.10cxx11abiTRUE-cp313-cp313-linux_x86_64.whl" in url
+
     def test_exact_wheel_url_uses_full_env_tuple(self):
         url = ips._build_flash_attn_wheel_url(
             {
@@ -333,83 +281,22 @@ class TestEnsureFlashAttn:
         mock_probe.assert_not_called()
         mock_install_wheel.assert_not_called()
 
-    def test_blackwell_gpu_skips_install_with_warning(self):
-        step_messages: list[tuple[str, str]] = []
-
-        def fake_step(
-            label: str,
-            value: str,
-            color_fn = None,
-        ):
-            step_messages.append((label, value))
-
-        with (
-            mock.patch.object(ips, "NO_TORCH", False),
-            mock.patch.object(ips, "IS_WINDOWS", False),
-            mock.patch.object(ips, "IS_MACOS", False),
-            mock.patch.object(ips, "has_blackwell_gpu", return_value = True),
-            mock.patch.object(ips, "probe_torch_wheel_env") as mock_probe,
-            mock.patch.object(ips, "install_wheel") as mock_install_wheel,
-            mock.patch.object(ips, "_step", side_effect = fake_step),
-            mock.patch("subprocess.run", return_value = self._import_check()),
-        ):
-            ips._ensure_flash_attn()
-
-        mock_probe.assert_not_called()
-        mock_install_wheel.assert_not_called()
-        assert any(label == "warning" and "Blackwell" in msg for label, msg in step_messages)
-
-    def test_blackwell_gpu_on_windows_emits_blackwell_warning(self):
-        step_messages: list[tuple[str, str]] = []
-
-        def fake_step(
-            label: str,
-            value: str,
-            color_fn = None,
-        ):
-            step_messages.append((label, value))
-
+    def test_windows_skips_install_without_probing(self):
+        # flash-attn is Linux-only: on Windows the installer returns before
+        # probing the torch env or resolving a wheel (no Windows wheels are
+        # published upstream).
         with (
             mock.patch.object(ips, "NO_TORCH", False),
             mock.patch.object(ips, "IS_WINDOWS", True),
             mock.patch.object(ips, "IS_MACOS", False),
-            mock.patch.object(ips, "has_blackwell_gpu", return_value = True),
             mock.patch.object(ips, "probe_torch_wheel_env") as mock_probe,
             mock.patch.object(ips, "install_wheel") as mock_install_wheel,
-            mock.patch.object(ips, "_step", side_effect = fake_step),
             mock.patch("subprocess.run", return_value = self._import_check()),
         ):
             ips._ensure_flash_attn()
 
         mock_probe.assert_not_called()
         mock_install_wheel.assert_not_called()
-        assert any(label == "warning" and "Blackwell" in msg for label, msg in step_messages)
-
-    def test_non_blackwell_windows_does_not_emit_blackwell_warning(self):
-        step_messages: list[tuple[str, str]] = []
-
-        def fake_step(
-            label: str,
-            value: str,
-            color_fn = None,
-        ):
-            step_messages.append((label, value))
-
-        with (
-            mock.patch.object(ips, "NO_TORCH", False),
-            mock.patch.object(ips, "IS_WINDOWS", True),
-            mock.patch.object(ips, "IS_MACOS", False),
-            mock.patch.object(ips, "has_blackwell_gpu", return_value = False),
-            mock.patch.object(ips, "probe_torch_wheel_env") as mock_probe,
-            mock.patch.object(ips, "install_wheel") as mock_install_wheel,
-            mock.patch.object(ips, "_step", side_effect = fake_step),
-            mock.patch("subprocess.run", return_value = self._import_check()),
-        ):
-            ips._ensure_flash_attn()
-
-        mock_probe.assert_not_called()
-        mock_install_wheel.assert_not_called()
-        assert not any("Blackwell" in msg for _, msg in step_messages)
 
 
 class TestInstallPythonStackFlashAttnIntegration:
diff --git a/tests/python/test_mlx_public_trainer_api.py b/tests/python/test_mlx_public_trainer_api.py
index 2c33f86af1..89f304c76d 100644
--- a/tests/python/test_mlx_public_trainer_api.py
+++ b/tests/python/test_mlx_public_trainer_api.py
@@ -115,6 +115,9 @@ def test_mlx_training_arguments_accept_trl_style_kwargs():
 def test_mlx_training_arguments_do_not_warn_for_implemented_or_falsey_extras():
     """Implemented and falsey inert compatibility kwargs should stay quiet."""
     unsloth = _import_mlx_unsloth()
+    supported_eval_kwargs = {}
+    if "eval_strategy" in unsloth._MLX_TRAINING_CONFIG_FIELDS:
+        supported_eval_kwargs = {"eval_strategy": "no", "eval_delay": 1}
 
     with warnings.catch_warnings(record = True) as caught:
         warnings.simplefilter("always")
@@ -125,12 +128,16 @@ def test_mlx_training_arguments_do_not_warn_for_implemented_or_falsey_extras():
             remove_unused_columns = False,
             assistant_only_loss = False,
             completion_only_loss = False,
+            **supported_eval_kwargs,
         )
 
     assert args.warmup_steps == 2
     assert args.padding_free is False
     assert args.remove_unused_columns is False
     assert args.completion_only_loss is False
+    if supported_eval_kwargs:
+        assert args.eval_strategy == "no"
+        assert args.eval_delay == 1
     assert caught == []
 
 
@@ -705,17 +712,10 @@ def test_mlx_trainer_rejects_unsafe_unsupported_sft_kwargs():
         )
 
 
-def test_mlx_trainer_rejects_metrics_and_callbacks():
-    """Trainer hooks should fail because MLXTrainer cannot honor them yet."""
+def test_mlx_trainer_rejects_compute_metrics():
+    """compute_metrics is still unsupported by MLXTrainer."""
     unsloth = _import_mlx_unsloth()
 
-    with pytest.raises(NotImplementedError, match = "callbacks"):
-        unsloth.UnslothTrainer(
-            model = _DummyModel(),
-            tokenizer = None,
-            train_dataset = [],
-            callbacks = [object()],
-        )
     with pytest.raises(NotImplementedError, match = "compute_metrics"):
         unsloth.UnslothTrainer(
             model = _DummyModel(),
@@ -725,6 +725,46 @@ def test_mlx_trainer_rejects_metrics_and_callbacks():
         )
 
 
+def test_mlx_trainer_accepts_callbacks():
+    """Callbacks are routed to MLXTrainer when the zoo backend supports them."""
+    unsloth = _import_mlx_unsloth()
+    from transformers import TrainerCallback
+
+    if not unsloth._mlx_trainer_supports_kwarg("callbacks"):
+        pytest.skip("requires unsloth-zoo MLXTrainer callback support")
+
+    class Callback(TrainerCallback):
+        pass
+
+    trainer = unsloth.UnslothTrainer(
+        model = _DummyModel(),
+        tokenizer = None,
+        train_dataset = [],
+        callbacks = [Callback()],
+    )
+    assert any(isinstance(cb, Callback) for cb in trainer.callback_handler.callbacks)
+
+
+def test_mlx_trainer_rejects_callbacks_with_old_zoo(monkeypatch):
+    """Older unsloth-zoo builds should fail clearly instead of TypeError."""
+    unsloth = _import_mlx_unsloth()
+    from transformers import TrainerCallback
+
+    monkeypatch.setattr(
+        unsloth,
+        "_mlx_trainer_supports_kwarg",
+        lambda name: name != "callbacks",
+    )
+
+    with pytest.raises(NotImplementedError, match = "callbacks require"):
+        unsloth.UnslothTrainer(
+            model = _DummyModel(),
+            tokenizer = None,
+            train_dataset = [],
+            callbacks = [TrainerCallback()],
+        )
+
+
 def test_mlx_trainer_rejects_custom_data_collator():
     """MLXTrainer owns batching; custom SFT data collators must not be ignored."""
     unsloth = _import_mlx_unsloth()
diff --git a/tests/python/test_remove_special_tokens_no_bos.py b/tests/python/test_remove_special_tokens_no_bos.py
new file mode 100644
index 0000000000..94c5ea3027
--- /dev/null
+++ b/tests/python/test_remove_special_tokens_no_bos.py
@@ -0,0 +1,46 @@
+import ast
+from pathlib import Path
+
+
+def _load_remove_special_tokens():
+    # Extract remove_special_tokens without importing unsloth (importing unsloth
+    # needs unsloth_zoo / a GPU). The function is pure Python and uses no imports,
+    # so it execs cleanly in an empty namespace.
+    source = Path(__file__).parents[2] / "unsloth" / "chat_templates.py"
+    tree = ast.parse(source.read_text(encoding = "utf-8"))
+    funcs = [
+        node
+        for node in tree.body
+        if isinstance(node, ast.FunctionDef) and node.name == "remove_special_tokens"
+    ]
+    namespace = {}
+    module = ast.Module(body = funcs, type_ignores = [])
+    ast.fix_missing_locations(module)
+    exec(compile(module, str(source), "exec"), namespace)
+    return namespace["remove_special_tokens"]
+
+
+class _StubTokenizer:
+    def __init__(self, bos_token):
+        self.bos_token = bos_token
+
+
+def test_no_bos_tokenizer_does_not_crash():
+    # Tokenizers such as Qwen2 / Qwen2.5, GPT-2, Falcon and GPT-NeoX have no BOS
+    # token, so tokenizer.bos_token is None. remove_special_tokens must leave the
+    # prompt untouched instead of raising
+    # "TypeError: startswith first arg must be str or a tuple of str, not NoneType".
+    remove_special_tokens = _load_remove_special_tokens()
+    assert remove_special_tokens(_StubTokenizer(None), "Hello world") == "Hello world"
+
+
+def test_double_bos_is_stripped():
+    # A tokenizer with a BOS token still has a single leading BOS removed.
+    remove_special_tokens = _load_remove_special_tokens()
+    assert remove_special_tokens(_StubTokenizer("<s>"), "<s>Hello world") == "Hello world"
+
+
+def test_prompt_without_leading_bos_unchanged():
+    # A BOS-bearing tokenizer leaves a prompt that does not start with BOS alone.
+    remove_special_tokens = _load_remove_special_tokens()
+    assert remove_special_tokens(_StubTokenizer("<s>"), "Hello world") == "Hello world"
diff --git a/tests/python/test_to_sharegpt_optional_none.py b/tests/python/test_to_sharegpt_optional_none.py
new file mode 100644
index 0000000000..05fcb8a6a8
--- /dev/null
+++ b/tests/python/test_to_sharegpt_optional_none.py
@@ -0,0 +1,97 @@
+import ast
+import re
+from pathlib import Path
+
+
+def _load_formatter_builders():
+    # Extract _parse_combined_prompt and _create_formatter without importing
+    # unsloth (importing unsloth needs unsloth_zoo / a GPU). Both are pure
+    # Python and only use the `re` module.
+    source = Path(__file__).parents[2] / "unsloth" / "chat_templates.py"
+    tree = ast.parse(source.read_text(encoding = "utf-8"))
+    wanted = {"_parse_combined_prompt", "_create_formatter"}
+    funcs = [
+        node for node in tree.body if isinstance(node, ast.FunctionDef) and node.name in wanted
+    ]
+    namespace = {"re": re}
+    module = ast.Module(body = funcs, type_ignores = [])
+    ast.fix_missing_locations(module)
+    exec(compile(module, str(source), "exec"), namespace)
+    return namespace["_parse_combined_prompt"], namespace["_create_formatter"]
+
+
+class _StubDataset:
+    def __init__(self, column_names):
+        self.column_names = column_names
+
+
+def _render(merged_prompt, columns, batch):
+    parse, create = _load_formatter_builders()
+    possible_columns, final_optional_prompts = parse(merged_prompt, _StubDataset(columns))
+    processor = create(possible_columns, final_optional_prompts, "text")
+    return processor(batch)["text"]
+
+
+def test_optional_block_missing_second_column_does_not_render_none():
+    # A [[...]] block may reference several columns; only the first gates the
+    # block. A later column that is None must not render as the literal "None".
+    merged_prompt = "Location: [[{city}, {country}]] end"
+    out = _render(
+        merged_prompt,
+        ["city", "country"],
+        {"city": ["Paris"], "country": [None]},
+    )
+    assert out[0] == "Location: Paris,  end"
+    assert "None" not in out[0]
+
+
+def test_optional_block_all_columns_present_unchanged():
+    merged_prompt = "Location: [[{city}, {country}]] end"
+    out = _render(
+        merged_prompt,
+        ["city", "country"],
+        {"city": ["Paris"], "country": ["France"]},
+    )
+    assert out[0] == "Location: Paris, France end"
+
+
+def test_optional_block_gating_column_empty_is_dropped():
+    # When the gating (first) column is empty the whole block is omitted; this
+    # behaviour is unchanged by the None coercion.
+    merged_prompt = "Location: [[{city}, {country}]] end"
+    out = _render(
+        merged_prompt,
+        ["city", "country"],
+        {"city": [""], "country": ["France"]},
+    )
+    assert out[0] == "Location:  end"
+
+
+def test_single_column_optional_block_gated_out_on_none():
+    # Single-column blocks were already gated correctly (the sole column is the
+    # gate); confirm they stay unaffected.
+    merged_prompt = "Name: [[{name}]]!"
+    out = _render(merged_prompt, ["name"], {"name": [None, "Bob"]})
+    assert out == ["Name: !", "Name: Bob!"]
+
+
+def test_required_column_none_does_not_render_none():
+    # A required (non-[[...]]) column that is None must not render as the
+    # literal "None" either; coercion happens at the row source, so both the
+    # required and optional branches are covered.
+    merged_prompt = "Location: {city}, {country} end"
+    out = _render(
+        merged_prompt,
+        ["city", "country"],
+        {"city": ["Paris"], "country": [None]},
+    )
+    assert out[0] == "Location: Paris,  end"
+    assert "None" not in out[0]
+
+
+def test_optional_block_falsy_but_present_gating_value_still_renders():
+    # The gate keeps a block whenever the first column is not "". A falsy but
+    # real value (0) must not be treated as absent, so the block still renders.
+    merged_prompt = "Count: [[{n}]]!"
+    out = _render(merged_prompt, ["n"], {"n": [0]})
+    assert out[0] == "Count: 0!"
diff --git a/tests/python/test_tokenizers_and_torch_constraint.py b/tests/python/test_tokenizers_and_torch_constraint.py
index 7390d7be9b..4322f0c7d6 100644
--- a/tests/python/test_tokenizers_and_torch_constraint.py
+++ b/tests/python/test_tokenizers_and_torch_constraint.py
@@ -14,6 +14,7 @@ _TESTS_DIR = pathlib.Path(__file__).resolve().parent.parent  # tests/
 _REPO_ROOT = _TESTS_DIR.parent  # unsloth/
 _INSTALL_SH = _REPO_ROOT / "install.sh"
 _INSTALL_PS1 = _REPO_ROOT / "install.ps1"
+_SETUP_PS1 = _REPO_ROOT / "studio" / "setup.ps1"
 _NO_TORCH_RT = _REPO_ROOT / "studio" / "backend" / "requirements" / "no-torch-runtime.txt"
 
 
@@ -109,6 +110,56 @@ class TestStructuralInstallPs1Unchanged:
         assert '"torch>=2.4,<2.11.0"' in self._ps1
 
 
+class TestInstallPs1UvDefaultIndex:
+    """Installer-managed torch indexes must override inherited uv defaults."""
+
+    _ps1 = _read(_INSTALL_PS1)
+
+    def test_torch_installs_use_default_index(self):
+        assert "--default-index $TorchIndexUrl" in self._ps1
+        assert "--default-index $ROCmIndexUrl" in self._ps1
+
+    def test_torch_installs_do_not_use_deprecated_index_url(self):
+        assert "--index-url $TorchIndexUrl" not in self._ps1
+        assert "--index-url $ROCmIndexUrl" not in self._ps1
+
+    def test_torch_installs_neutralize_all_uv_index_env_vars(self):
+        # Extra-index vars outrank --default-index, so pinned installs must clear them.
+        for var in ("UV_DEFAULT_INDEX", "UV_INDEX_URL", "UV_INDEX", "UV_EXTRA_INDEX_URL"):
+            assert var in self._ps1
+        assert 'Remove-Item "Env:$n"' in self._ps1
+
+
+class TestSetupPs1FastInstallIndex:
+    """setup.ps1 Fast-Install must neutralize inherited uv indexes when pinning."""
+
+    _ps1 = _read(_SETUP_PS1)
+
+    def test_fast_install_clears_all_uv_index_env_vars(self):
+        for var in ("UV_DEFAULT_INDEX", "UV_INDEX_URL", "UV_INDEX", "UV_EXTRA_INDEX_URL"):
+            assert var in self._ps1
+        # Must truly remove the vars (child sees no value), not set them empty.
+        assert 'Remove-Item "Env:$n"' in self._ps1
+
+
+class TestInstallShUvDefaultIndex:
+    """Linux/Mac installer torch indexes must override inherited uv defaults."""
+
+    _sh = _read(_INSTALL_SH)
+
+    def test_torch_installs_use_default_index(self):
+        assert '--default-index "$TORCH_INDEX_URL"' in self._sh
+
+    def test_torch_installs_do_not_use_deprecated_index_url(self):
+        assert '--index-url "$TORCH_INDEX_URL"' not in self._sh
+
+    def test_torch_installs_neutralize_all_uv_index_env_vars(self):
+        # --default-index installs run with all uv index env vars unset via `env -u`.
+        assert (
+            "env -u UV_DEFAULT_INDEX -u UV_INDEX_URL -u UV_INDEX -u UV_EXTRA_INDEX_URL" in self._sh
+        )
+
+
 # Group 2 -- Shell snippet tests (bash subprocess, mocked python)
 class TestTorchConstraintShell:
     """Test the TORCH_CONSTRAINT block via bash with mocked python minor versions."""
diff --git a/tests/sh/test_mac_intel_compat.sh b/tests/sh/test_mac_intel_compat.sh
index 3c3bbfaa5f..8a0ff4b641 100644
--- a/tests/sh/test_mac_intel_compat.sh
+++ b/tests/sh/test_mac_intel_compat.sh
@@ -312,7 +312,7 @@ if [ "$SKIP_TORCH" = true ]; then
 else
     echo "==> Installing PyTorch ($TORCH_INDEX_URL)..."
     uv pip install --python "$_VENV_PY" "torch>=2.4,<2.11.0" torchvision torchaudio \
-        --index-url "$TORCH_INDEX_URL"
+        --default-index "$TORCH_INDEX_URL"
 fi
 TORCH_EOF
 
diff --git a/tests/studio/install/test_llama_pr_force_and_source.py b/tests/studio/install/test_llama_pr_force_and_source.py
index 8f40c9d720..4ff8c349c3 100644
--- a/tests/studio/install/test_llama_pr_force_and_source.py
+++ b/tests/studio/install/test_llama_pr_force_and_source.py
@@ -362,8 +362,11 @@ class TestSourcePatternsSh:
         assert '_LLAMA_SOURCE="${_DEFAULT_LLAMA_SOURCE}"' in self.content
 
     def test_release_repo_override_removed(self):
+        # No env-based release-repo override, and CPU-only hosts no longer fall
+        # back to ggml-org -- every host now routes to the fork.
         assert "UNSLOTH_LLAMA_RELEASE_REPO:-unslothai/llama.cpp" not in self.content
-        assert '_HELPER_RELEASE_REPO="ggml-org/llama.cpp"' in self.content
+        assert '_HELPER_RELEASE_REPO="unslothai/llama.cpp"' in self.content
+        assert '_HELPER_RELEASE_REPO="ggml-org/llama.cpp"' not in self.content
 
     def test_force_compile_skips_prebuilt_resolution_early(self):
         assert 'if [ "$_LLAMA_FORCE_COMPILE" = "1" ]; then' in self.content
@@ -425,12 +428,11 @@ class TestSourcePatternsPs1:
         assert "$LlamaSource = $DefaultLlamaSource" in self.content
 
     def test_release_repo_override_removed(self):
-        # Repo chosen by GPU detection (GPU -> fork, CPU -> ggml-org), no env override.
+        # No env-based release-repo override; every host now routes to the fork
+        # (the CPU-only ggml-org fallback was removed), mirroring setup.sh.
         assert "$HelperReleaseRepo = if ($env:UNSLOTH_LLAMA_RELEASE_REPO)" not in self.content
-        assert (
-            "$HelperReleaseRepo = if ($HasNvidiaSmi -or $HasROCm -or $script:ROCmGfxArch) "
-            '{ "unslothai/llama.cpp" } else { "ggml-org/llama.cpp" }' in self.content
-        )
+        assert '$HelperReleaseRepo = "unslothai/llama.cpp"' in self.content
+        assert "$HelperReleaseRepo = if (" not in self.content
 
     def test_force_compile_skips_prebuilt_resolution_early(self):
         assert 'if ($env:UNSLOTH_LLAMA_FORCE_COMPILE -eq "1") {' in self.content
diff --git a/tests/studio/install/test_pr4562_bugfixes.py b/tests/studio/install/test_pr4562_bugfixes.py
index a8cec0cb85..0d2b092924 100644
--- a/tests/studio/install/test_pr4562_bugfixes.py
+++ b/tests/studio/install/test_pr4562_bugfixes.py
@@ -658,15 +658,21 @@ class TestSourceCodePatterns:
         assert "_HELPER_RELEASE_REPO}/releases/latest" not in content
         assert "ggml-org/llama.cpp/releases/latest" not in content
 
-    def test_setup_sh_routes_to_fork_only_on_usable_gpu(self):
-        """Linux routing gates NVIDIA on GPU usability, not nvidia-smi presence, so
-        CPU-only/hidden-GPU hosts get the ggml CPU prebuilt. Guards the old presence-only loop."""
+    def test_setup_sh_routes_every_host_to_fork(self):
+        """CPU-only Linux (the last ggml-org artifact consumer) now routes to the
+        fork like every other host, so the release-repo decision is unconditional.
+        Guards against a silent reintroduction of a ggml-org CPU routing branch.
+        GPU usability detection (used for PyTorch / source decisions) must stay."""
         content = SETUP_SH.read_text()
+        assert '_HELPER_RELEASE_REPO="unslothai/llama.cpp"' in content
+        assert '_HELPER_RELEASE_REPO="ggml-org/llama.cpp"' not in content
+        # Usability gating (not routing) still distinguishes a hidden GPU.
         assert '[ "$_setup_nvidia_usable" = true ]' in content
         assert "CUDA_VISIBLE_DEVICES" in content
-        # nvidia-smi must NOT be back in the bare presence loop.
-        assert "for _GPU_TOOL in nvidia-smi" not in content
-        assert "for _GPU_TOOL in rocminfo amd-smi hipconfig hipinfo" in content
+        # The GPU-tooling probe (PR #4562) stays: ROCm detection goes through
+        # command -v, not a bare presence loop that mishandled a hidden nvidia-smi.
+        assert "command -v rocminfo" in content
+        assert "command -v amd-smi" in content
 
     def test_setup_sh_reports_installed_prebuilt_release(self):
         """Shell wrapper should report the installed prebuilt release from metadata."""
diff --git a/tests/studio/install/test_rocm_support.py b/tests/studio/install/test_rocm_support.py
index bfc8132683..5cabf41f57 100644
--- a/tests/studio/install/test_rocm_support.py
+++ b/tests/studio/install/test_rocm_support.py
@@ -3164,21 +3164,25 @@ class TestRocmGfxForwarding:
         assert "--rocm-gfx" in source
         assert "$script:ROCmGfxArch" in source
 
-    def test_setup_sh_routes_inferred_gfx_to_fork(self):
-        # An inferred gfx arch must route to the fork even without ROCm tooling.
-        # Pin the specific guard (a bare "${_setup_gfx:-}" also appears elsewhere).
+    def test_setup_sh_routes_unconditionally_to_fork(self):
+        # CPU-only hosts no longer fall back to ggml-org -- the release-repo
+        # decision is an unconditional fork assignment now. Pin the line text.
         source = _SETUP_SH_PATH.read_text(encoding = "utf-8")
-        assert '[ "$_LINUX_HAS_GPU" = false ] && [ -n "${_setup_gfx:-}" ]' in source
+        assert '_HELPER_RELEASE_REPO="unslothai/llama.cpp"' in source
+        assert '_HELPER_RELEASE_REPO="ggml-org/llama.cpp"' not in source
 
-    def test_setup_ps1_routes_inferred_gfx_to_fork(self):
-        # On Windows, a resolved $script:ROCmGfxArch counts as a fork install
-        # even when $HasROCm is false (Adrenalin-only, no HIP runtime).
+    def test_setup_ps1_routes_unconditionally_to_fork(self):
+        # Same on Windows: the fork now ships the windows-cpu / windows-arm64
+        # bundles, so $HelperReleaseRepo is an unconditional fork assignment.
         source = _SETUP_PS1_PATH.read_text(encoding = "utf-8")
-        assert "$HasNvidiaSmi -or $HasROCm -or $script:ROCmGfxArch" in source
+        assert '$HelperReleaseRepo = "unslothai/llama.cpp"' in source
+        assert "$HelperReleaseRepo = if (" not in source
 
-    # The assertions above pin the guard *text*; the tests below *execute* the
-    # real routing block and assert the resolved repo, so a refactor that keeps
-    # the literal but breaks the inferred-gfx -> fork decision is still caught.
+    # The text pins above guard the literal. The tests below *execute* the real
+    # routing line from setup.sh / setup.ps1 and assert the resolved release repo,
+    # so a refactor that reintroduces a conditional (or a ggml-org branch) is still
+    # caught. Inputs are varied -- CPU-only, inferred/forwarded gfx, usable NVIDIA --
+    # to prove no host slips back onto ggml-org. No GPU, no tooling, no network.
 
     @staticmethod
     def _resolve_setup_sh_repo(
@@ -3187,15 +3191,18 @@ class TestRocmGfxForwarding:
         setup_gfx,
         rocm_gfx_arch_env = "",
     ):
-        """Run setup.sh's routing block under bash with PATH emptied (no ROCm tooling) and return _HELPER_RELEASE_REPO."""
+        """Run setup.sh's release-repo routing block under bash and return the
+        resolved _HELPER_RELEASE_REPO. PATH is emptied so any stray tooling probe
+        misses; routing is unconditional, so the GPU inputs only prove no branch
+        reroutes a host to ggml-org."""
         import shutil
 
         bash = shutil.which("bash")
         if bash is None:
             pytest.skip("bash not available")
         source = _SETUP_SH_PATH.read_text(encoding = "utf-8")
-        start = source.index("\n_LINUX_HAS_GPU=false\n") + 1
-        end = source.index("\nunset _GPU_TOOL", start) + len("\nunset _GPU_TOOL")
+        start = source.index('\n_HELPER_RELEASE_REPO="unslothai/llama.cpp"\n') + 1
+        end = source.index("\n_LLAMA_PR=", start)
         block = source[start:end]
         assert "_HELPER_RELEASE_REPO" in block, "setup.sh routing anchors not found"
         env = {
@@ -3217,25 +3224,31 @@ class TestRocmGfxForwarding:
         assert result.returncode == 0, result.stderr
         return result.stdout.strip()
 
-    def test_setup_sh_inferred_gfx_resolves_to_fork(self):
-        # Only a name-inferred gfx arch -> route to the fork's per-gfx prebuilt
-        # (not ggml-org). x64 and arm64 share the fork branch.
-        assert self._resolve_setup_sh_repo("x86_64", False, "gfx1100") == "unslothai/llama.cpp"
-        assert self._resolve_setup_sh_repo("aarch64", False, "gfx1100") == "unslothai/llama.cpp"
-
-    def test_setup_sh_env_forwarded_gfx_resolves_to_fork(self):
-        # No probe fired but UNSLOTH_ROCM_GFX_ARCH is set: adopt the env arch
-        # and route to the fork, same as name-inference.
-        repo = self._resolve_setup_sh_repo("x86_64", False, "", rocm_gfx_arch_env = "gfx1100")
-        assert repo == "unslothai/llama.cpp"
-
-    def test_setup_sh_cpu_host_still_resolves_to_ggml(self):
-        # A real CPU host (no GPU, no inferred gfx, no env override) must keep routing to ggml-org.
-        assert self._resolve_setup_sh_repo("x86_64", False, "") == "ggml-org/llama.cpp"
+    @pytest.mark.parametrize(
+        "machine, nvidia_usable, setup_gfx, env_gfx",
+        [
+            ("x86_64", False, "", ""),  # plain CPU host (used to take ggml-org)
+            ("aarch64", False, "", ""),  # plain CPU arm64 host (used to take ggml-org)
+            ("x86_64", False, "gfx1100", ""),  # name-inferred gfx
+            ("x86_64", False, "", "gfx1100"),  # env-forwarded gfx
+            ("x86_64", True, "", ""),  # usable NVIDIA
+        ],
+    )
+    def test_setup_sh_routing_block_always_resolves_to_fork(
+        self, machine, nvidia_usable, setup_gfx, env_gfx
+    ):
+        assert (
+            self._resolve_setup_sh_repo(
+                machine, nvidia_usable, setup_gfx, rocm_gfx_arch_env = env_gfx
+            )
+            == "unslothai/llama.cpp"
+        )
 
     @staticmethod
-    def _resolve_setup_ps1_repo(has_nvidia, has_rocm, gfx_arch):
-        """Run setup.ps1's $HelperReleaseRepo selection under pwsh and return the resolved repo."""
+    def _resolve_setup_ps1_repo():
+        """Run setup.ps1's $HelperReleaseRepo assignment under pwsh and return the
+        resolved repo. The assignment is unconditional now, so there are no host
+        inputs to vary."""
         import shutil
 
         pwsh = shutil.which("pwsh")
@@ -3243,21 +3256,11 @@ class TestRocmGfxForwarding:
             pytest.skip("pwsh not available")
         source = _SETUP_PS1_PATH.read_text(encoding = "utf-8")
         line = next(
-            (
-                ln
-                for ln in source.splitlines()
-                if ln.strip().startswith("$HelperReleaseRepo = if (")
-            ),
+            (ln for ln in source.splitlines() if ln.strip().startswith("$HelperReleaseRepo =")),
             None,
         )
         assert line is not None, "$HelperReleaseRepo selection not found in setup.ps1"
-        harness = (
-            f"$HasNvidiaSmi = ${'true' if has_nvidia else 'false'}\n"
-            f"$HasROCm = ${'true' if has_rocm else 'false'}\n"
-            f"$script:ROCmGfxArch = '{gfx_arch}'\n"
-            f"{line}\n"
-            "Write-Output $HelperReleaseRepo"
-        )
+        harness = f"{line}\nWrite-Output $HelperReleaseRepo"
         result = subprocess.run(
             [pwsh, "-NoProfile", "-Command", harness],
             capture_output = True,
@@ -3267,13 +3270,10 @@ class TestRocmGfxForwarding:
         assert result.returncode == 0, result.stderr
         return result.stdout.strip()
 
-    def test_setup_ps1_inferred_gfx_resolves_to_fork(self):
-        # Adrenalin-only host: $HasROCm false but gfx inferred -> fork's windows-rocm bundle.
-        assert self._resolve_setup_ps1_repo(False, False, "gfx1100") == "unslothai/llama.cpp"
-
-    def test_setup_ps1_cpu_host_still_resolves_to_ggml(self):
-        # No NVIDIA, no ROCm, no inferred gfx -> CPU host stays on ggml-org.
-        assert self._resolve_setup_ps1_repo(False, False, "") == "ggml-org/llama.cpp"
+    def test_setup_ps1_routing_resolves_to_fork(self):
+        # Windows routing is unconditional now: CPU-only Windows (x64 and arm64)
+        # uses the fork's windows-cpu / windows-arm64 bundles, not ggml-org.
+        assert self._resolve_setup_ps1_repo() == "unslothai/llama.cpp"
 
 
 # TEST: _pick_rocm_gfx_target -- visible-device selection from rocminfo output.
diff --git a/tests/studio/install/test_selection_logic.py b/tests/studio/install/test_selection_logic.py
index 464aa9a578..7d3120adb8 100644
--- a/tests/studio/install/test_selection_logic.py
+++ b/tests/studio/install/test_selection_logic.py
@@ -1720,8 +1720,10 @@ class TestResolveInstallAttempts:
         assert approved.release_tag == "llama-prebuilt-latest"
 
     def test_linux_cpu_fork_without_bundle_raises_no_upstream_fallback(self, monkeypatch):
-        # CPU-only Linux on the fork must not fall back to the ggml-org CPU asset; with
-        # no fork CPU bundle the resolver raises rather than reaching upstream.
+        # A CPU-only Linux host on the fork never falls back to the ggml-org CPU
+        # asset. CPU-only Linux now routes to the fork, but if a release manifest
+        # happens to ship no CPU bundle the resolver raises rather than quietly
+        # reaching for an upstream asset.
         host = make_host(
             has_usable_nvidia = False,
             has_physical_nvidia = False,
@@ -1846,6 +1848,151 @@ class TestResolveInstallAttempts:
         assert attempts[0].name == asset_name
         assert attempts[0].source_label == "published"
 
+    @pytest.mark.parametrize(
+        "system, machine, asset_name, install_kind, bundle_profile",
+        [
+            # CPU-only Linux x64 -> fork linux-cpu (was ggml-org ubuntu-x64).
+            ("Linux", "x86_64", "app-b9625-linux-x64-cpu.tar.gz", "linux-cpu", "linux-cpu-x64"),
+            # CPU-only Linux arm64 -> fork linux-arm64 (was ggml-org ubuntu-arm64).
+            (
+                "Linux",
+                "aarch64",
+                "app-b9625-linux-arm64-cpu.tar.gz",
+                "linux-arm64",
+                "linux-cpu-arm64",
+            ),
+            # CPU-only Windows arm64 -> fork windows-arm64 (was ggml-org win-cpu-arm64).
+            (
+                "Windows",
+                "arm64",
+                "app-b9625-windows-arm64-cpu.zip",
+                "windows-arm64",
+                "windows-cpu-arm64",
+            ),
+        ],
+    )
+    def test_cpu_host_prefers_published_fork_asset(
+        self, monkeypatch, system, machine, asset_name, install_kind, bundle_profile
+    ):
+        # CPU-only hosts now select the fork's CPU bundle from the manifest and
+        # must never query ggml-org upstream assets. Windows x64 CPU is covered
+        # separately by test_windows_cpu_prefers_published_asset.
+        host = make_host(
+            system = system,
+            machine = machine,
+            has_usable_nvidia = False,
+            has_physical_nvidia = False,
+            nvidia_smi = None,
+        )
+        release = make_release(
+            [
+                make_artifact(
+                    asset_name,
+                    install_kind = install_kind,
+                    runtime_line = None,
+                    coverage_class = None,
+                    supported_sms = [],
+                    min_sm = None,
+                    max_sm = None,
+                    bundle_profile = bundle_profile,
+                    rank = 1000,
+                )
+            ],
+            release_tag = "llama-prebuilt-latest",
+            upstream_tag = "b9625",
+            assets = {asset_name: f"https://published.example/{asset_name}"},
+        )
+        checksums = make_checksums_with_source(
+            [asset_name],
+            release_tag = release.release_tag,
+            upstream_tag = "b9625",
+        )
+
+        monkeypatch.setattr(
+            INSTALL_LLAMA_PREBUILT,
+            "iter_resolved_published_releases",
+            lambda requested_tag, published_repo, published_release_tag = "": iter(
+                [
+                    INSTALL_LLAMA_PREBUILT.ResolvedPublishedRelease(
+                        bundle = release,
+                        checksums = checksums,
+                    )
+                ]
+            ),
+        )
+        monkeypatch.setattr(
+            INSTALL_LLAMA_PREBUILT,
+            "github_release_assets",
+            lambda repo, tag: (_ for _ in ()).throw(
+                AssertionError("fork CPU host must not query upstream assets")
+            ),
+        )
+
+        _requested_tag, resolved_tag, attempts, _approved = resolve_install_attempts(
+            "latest",
+            host,
+            "unslothai/llama.cpp",
+            "",
+        )
+
+        assert resolved_tag == "b9625"
+        assert attempts[0].name == asset_name
+        assert attempts[0].install_kind == install_kind
+        assert attempts[0].source_label == "published"
+
+    def test_cpu_only_unsupported_arch_source_builds(self, monkeypatch):
+        # A CPU-only Linux host that is neither x86_64 nor arm64 (ppc64le,
+        # riscv64, s390x) has no compatible CPU bundle. It must source-build, not
+        # receive the x86_64 linux-cpu binary (the Linux preflight checks libs,
+        # not ELF arch, so a wrong-arch binary would slip through).
+        host = make_host(
+            machine = "ppc64le",
+            has_usable_nvidia = False,
+            has_physical_nvidia = False,
+            nvidia_smi = None,
+        )
+        assert not host.is_x86_64 and not host.is_arm64
+        x64_asset = "app-b9625-linux-x64-cpu.tar.gz"
+        release = make_release(
+            [
+                make_artifact(
+                    x64_asset,
+                    install_kind = "linux-cpu",
+                    runtime_line = None,
+                    coverage_class = None,
+                    supported_sms = [],
+                    min_sm = None,
+                    max_sm = None,
+                    bundle_profile = "linux-cpu-x64",
+                    rank = 1000,
+                )
+            ],
+            release_tag = "llama-prebuilt-latest",
+            upstream_tag = "b9625",
+            assets = {x64_asset: f"https://published.example/{x64_asset}"},
+        )
+        checksums = make_checksums_with_source(
+            [x64_asset],
+            release_tag = release.release_tag,
+            upstream_tag = "b9625",
+        )
+
+        monkeypatch.setattr(
+            INSTALL_LLAMA_PREBUILT,
+            "iter_resolved_published_releases",
+            lambda requested_tag, published_repo, published_release_tag = "": iter(
+                [
+                    INSTALL_LLAMA_PREBUILT.ResolvedPublishedRelease(
+                        bundle = release,
+                        checksums = checksums,
+                    )
+                ]
+            ),
+        )
+
+        with pytest.raises(PrebuiltFallback, match = "no compatible Linux prebuilt asset was found"):
+            resolve_install_attempts("latest", host, "unslothai/llama.cpp", "")
+
     def test_macos_prefers_published_asset(self, monkeypatch):
         host = make_host(
             system = "Darwin",
@@ -2537,68 +2684,6 @@ class TestBlackwellCuda124Exclusion:
         assert kept == [cpu]
 
 
-# N.1c3. direct_linux_release_plan -- no silent CPU on NVIDIA hosts
-
-
-class TestDirectLinuxNvidiaCpuGate:
-    """A linux-cpu-only release on an NVIDIA host must raise (caller walks back to a usable CUDA line), not silently CPU-install. CPU-only hosts keep the CPU bundle."""
-
-    def _bundle_cpu_only(self):
-        return make_release(
-            [
-                make_artifact(
-                    "llama-b8508-bin-ubuntu-x64.tar.gz",
-                    install_kind = "linux-cpu",
-                    runtime_line = None,
-                    coverage_class = None,
-                    supported_sms = [],
-                    min_sm = None,
-                    max_sm = None,
-                    bundle_profile = None,
-                ),
-            ]
-        )
-
-    def _patch(self, monkeypatch):
-        monkeypatch.setattr(
-            INSTALL_LLAMA_PREBUILT,
-            "parse_direct_linux_release_bundle",
-            lambda repo, release: self._bundle_cpu_only(),
-        )
-        monkeypatch.setattr(
-            INSTALL_LLAMA_PREBUILT,
-            "detect_torch_cuda_runtime_preference",
-            lambda host: CudaRuntimePreference(runtime_line = None, selection_log = []),
-        )
-        monkeypatch.setattr(
-            INSTALL_LLAMA_PREBUILT,
-            "detected_linux_runtime_lines",
-            lambda: (["cuda13"], {"cuda13": ["/usr/local/cuda/lib64"]}),
-        )
-
-    def test_nvidia_host_without_cuda_line_raises_for_walkback(self, monkeypatch):
-        self._patch(monkeypatch)
-        host = make_host(driver_cuda_version = (13, 1), compute_caps = ["100"])
-        with pytest.raises(PrebuiltFallback, match = "no compatible Linux prebuilt"):
-            INSTALL_LLAMA_PREBUILT.direct_linux_release_plan(
-                {"tag_name": "b8508"}, host, "unslothai/llama.cpp", "latest"
-            )
-
-    def test_cpu_host_still_gets_cpu_bundle(self, monkeypatch):
-        self._patch(monkeypatch)
-        host = make_host(
-            nvidia_smi = None,
-            driver_cuda_version = None,
-            compute_caps = [],
-            has_physical_nvidia = False,
-            has_usable_nvidia = False,
-        )
-        plan = INSTALL_LLAMA_PREBUILT.direct_linux_release_plan(
-            {"tag_name": "b8508"}, host, "unslothai/llama.cpp", "latest"
-        )
-        assert [a.install_kind for a in plan.attempts] == ["linux-cpu"]
-
-
 class TestLinuxPublishedAttemptsNvidiaCpuGate:
     """Live fork-manifest path: an NVIDIA host whose CUDA selection finds nothing gets an empty attempt list (source-builds with CUDA), not the manifest CPU bundle. CPU-only hosts still get the CPU bundle."""
 
@@ -3447,8 +3532,9 @@ class TestLinuxArm64ForkFallsBackToSource:
         assert plans == ["plan"]
 
     def test_arm64_cpu_on_ggml_org_is_not_blocked(self, monkeypatch):
-        # CPU-only arm64 routes to ggml-org, so the guard must not fire; it reaches the
-        # iterator (empty here -> generic message).
+        # ggml-org is reachable only via an explicit --published-repo override now,
+        # but the guard must still not fire on arm64 there; it reaches the iterator
+        # (empty here -> generic message).
         monkeypatch.setattr(
             INSTALL_LLAMA_PREBUILT,
             "iter_release_payloads_by_time",
@@ -3474,7 +3560,7 @@ class TestLinuxArm64ForkFallsBackToSource:
 
 
 class TestCpuFallback:
-    """--cpu-fallback drops GPU attributes so the host's OS/arch CPU prebuilt is selected, letting an arm64 GPU host install ggml-org's arm64 CPU build when its source build produced no binary."""
+    """--cpu-fallback drops GPU attributes so the host's OS/arch CPU prebuilt is selected, letting an arm64 GPU host install the fork's arm64 CPU bundle when its source build produced no binary."""
 
     _SETUP_SH = PACKAGE_ROOT / "studio" / "setup.sh"
 
@@ -3553,10 +3639,15 @@ class TestCpuFallback:
 
     def test_setup_sh_has_arm64_cpu_prebuilt_fallback(self):
         source = self._SETUP_SH.read_text(encoding = "utf-8")
-        assert "--cpu-fallback" in source
-        # Fallback targets ggml-org (only repo with an arm64 Linux build), gated on a
-        # degraded arm64 source build.
-        assert "ggml-org/llama.cpp" in source
+        # The arm64 GPU last-resort CPU fallback now pulls the fork's arm64 CPU
+        # bundle (app-<tag>-linux-arm64-cpu.tar.gz), not ggml-org's, and is gated
+        # on a degraded source build for arm64.
+        start = source.index("_ARM64_CPU_CMD=(")
+        end = source.index(")", start)
+        block = source[start:end]
+        assert "--cpu-fallback" in block
+        assert '--published-repo "unslothai/llama.cpp"' in block
+        assert '--published-repo "ggml-org/llama.cpp"' not in block
         assert "_LLAMA_CPP_DEGRADED" in source
 
 
diff --git a/tests/studio/playwright_chat_ui.py b/tests/studio/playwright_chat_ui.py
index a53534acc0..6a88b98c19 100644
--- a/tests/studio/playwright_chat_ui.py
+++ b/tests/studio/playwright_chat_ui.py
@@ -380,6 +380,18 @@ with sync_playwright() as p:
                 fail(f"/api/auth/refresh wedged: {refresh_resp['error']!r}")
             refresh = refresh_resp.get("body") or {}
             token = (refresh or {}).get("access_token")
+            next_refresh_token = (refresh or {}).get("refresh_token")
+            if token and next_refresh_token:
+                robust_evaluate(
+                    page,
+                    """([accessToken, refreshToken]) => {
+                        localStorage.setItem('unsloth_auth_token', accessToken);
+                        localStorage.setItem('unsloth_auth_refresh_token', refreshToken);
+                    }""",
+                    [token, next_refresh_token],
+                )
+            elif token:
+                fail("/api/auth/refresh returned access_token but no refresh_token")
     if not token:
         fail("could not obtain auth token after change-password")
 
@@ -1169,6 +1181,13 @@ with sync_playwright() as p:
         fail(f"curl login returned no access_token: {login_body!r}")
     info("CLI obtained an access token")
 
+    browser_refresh_token = robust_evaluate(
+        page,
+        "() => localStorage.getItem('unsloth_auth_refresh_token')",
+    )
+    if not browser_refresh_token:
+        fail("browser refresh token missing before CLI rotation")
+
     change_proc = subprocess.run(
         [
             "curl",
@@ -1203,18 +1222,39 @@ with sync_playwright() as p:
 
     # /change-password revoked refresh tokens server-side (auth.py), so
     # the browser's /api/auth/refresh must now fail.
-    refresh_after = evaluate_fetch(
-        page,
-        f"{BASE}/api/auth/refresh",
-        method = "POST",
-        timeout_ms = FETCH_TIMEOUT_MS,
+    refresh_proc = subprocess.run(
+        [
+            "curl",
+            "-sS",
+            "-o",
+            os.devnull,
+            "-w",
+            "%{http_code}",
+            "-X",
+            "POST",
+            f"{BASE}/api/auth/refresh",
+            "-H",
+            "Content-Type: application/json",
+            "-d",
+            json.dumps({"refresh_token": browser_refresh_token}),
+        ],
+        capture_output = True,
+        text = True,
+        timeout = 15,
     )
-    if refresh_after.get("error"):
-        fail(f"/api/auth/refresh wedged: {refresh_after['error']!r}")
-    if refresh_after["status"] == 200:
+    if refresh_proc.returncode != 0:
+        fail(
+            f"curl refresh-token check failed: rc={refresh_proc.returncode} "
+            f"stderr={refresh_proc.stderr!r} stdout={refresh_proc.stdout!r}"
+        )
+    try:
+        refresh_status = int(refresh_proc.stdout.strip())
+    except ValueError:
+        fail(f"curl refresh-token check returned invalid status: " f"{refresh_proc.stdout!r}")
+    if refresh_status == 200:
         fail(f"/api/auth/refresh should fail after CLI rotation; got 200")
     info(
-        f"OK browser /api/auth/refresh now {refresh_after['status']} "
+        f"OK browser /api/auth/refresh now {refresh_status} "
         "(refresh token revoked) -- old studio session can no longer renew"
     )
 
@@ -1224,22 +1264,124 @@ with sync_playwright() as p:
     # placeholder, and /api/health goes unreachable shortly after.
     # ─────────────────────────────────────────────────────
     step("Shutdown via account menu")
-    # Re-login with NEW2 for a valid /api/shutdown token (CLI rotation
-    # invalidated the old one). The stale token can make the SPA auth guard
-    # abort this goto with ERR_ABORTED, or redirect to the same /login URL
-    # ("interrupted by another navigation"); resolve on domcontentloaded and
-    # tolerate either -- the pw-field wait below confirms we are on /login.
-    _tolerated_nav = ("ERR_ABORTED", "interrupted by another navigation")
+    # Start fresh after the CLI rotation invalidates this browser session.
+    # Stay in the SAME context: macOS Chromium runs --single-process, where
+    # closing the last context kills the browser and a second context cannot
+    # be created. Open the new page before closing the old one; the context
+    # init script covers the new page.
     try:
-        page.goto(f"{BASE}/login", wait_until = "domcontentloaded", timeout = 60_000)
+        ctx.clear_cookies()
     except Exception as exc:
-        if not any(t in str(exc) for t in _tolerated_nav):
-            raise
-        info(f"goto /login interrupted ({exc!r}); password-field wait will confirm /login")
-    pw_field = page.locator("#password")
-    pw_field.wait_for(state = "visible", timeout = 60_000)
-    pw_field.fill(NEW2)
-    page.locator('button[type="submit"]').click()
+        info(f"WARN clearing stale session cookies failed: {exc!r}")
+    # Auth tokens live in localStorage, and /login's guest guard redirects on
+    # their mere presence, so drop them before navigating.
+    try:
+        page.evaluate(
+            "['unsloth_auth_token', 'unsloth_auth_refresh_token']"
+            ".forEach((key) => localStorage.removeItem(key))"
+        )
+    except Exception as exc:
+        info(f"WARN clearing stale auth tokens failed: {exc!r}")
+    _fresh_page = ctx.new_page()
+    _fresh_page.set_default_timeout(60_000)
+    _fresh_page.on("pageerror", lambda e: page_errors.append(str(e)))
+    _fresh_page.on("console", _on_console)
+    try:
+        page.close()
+    except Exception:
+        pass
+    page = _fresh_page
+
+    # Re-login with NEW2 for a valid /api/shutdown token. Route changes can
+    # still abort or interrupt this navigation, so the field wait below is the
+    # final confirmation that we reached /login.
+    _tolerated_nav = ("ERR_ABORTED", "interrupted by another navigation")
+    # A slow CI runner can make this re-login navigation time out even with the
+    # server healthy, so retry the whole goto/wait/fill/submit sequence (mirrors
+    # the change-password retry above). wait_for_health is a diagnostic pre-gate.
+    wait_for_health(BASE, timeout = 30.0, info = info)
+    relogin_err: Exception | None = None
+    for _relogin_attempt in range(3):
+        try:
+            try:
+                page.goto(f"{BASE}/login", wait_until = "domcontentloaded", timeout = 60_000)
+            except Exception as exc:
+                if not any(t in str(exc) for t in _tolerated_nav):
+                    raise
+                info(f"goto /login interrupted ({exc!r}); password-field wait will confirm /login")
+            pw_field = page.locator("#password")
+            pw_field.wait_for(state = "visible", timeout = 60_000)
+            pw_field.fill(NEW2)
+            # Wait on the login POST so a transient 4xx/5xx is caught and retried
+            # here, not swallowed until the out-of-loop composer wait.
+            status, _ = click_and_wait_for_response(
+                page,
+                url_substr = "/api/auth/login",
+                method = "POST",
+                do_click = lambda: page.locator('button[type="submit"]').click(),
+                timeout_ms = 30_000,
+                info = lambda m: print(f"[ui]   {m}", flush = True),
+            )
+            if status is not None and status >= 400:
+                raise AssertionError(
+                    f"login POST returned {status}; see console_errors={console_errors[:1]!r}"
+                )
+            relogin_err = None
+            break
+        except Exception as e:
+            relogin_err = e
+            try:
+                cur_url = page.url
+            except Exception:
+                cur_url = "<page closed>"
+            print(
+                f"[ui]   re-login attempt {_relogin_attempt + 1} failed: "
+                f"{type(e).__name__}: {str(e)[:200]}; page.url={cur_url}; "
+                f"page_errors={len(page_errors)} console_errors={len(console_errors)}",
+                flush = True,
+            )
+            if console_errors:
+                print(
+                    f"[ui]   first console.error: {console_errors[0][:200]!r}",
+                    flush = True,
+                )
+            if page_errors:
+                print(f"[ui]   first pageerror:    {page_errors[0][:200]!r}", flush = True)
+            try:
+                shoot(f"18-relogin-attempt-{_relogin_attempt + 1}-fail")
+            except Exception:
+                pass
+            if _relogin_attempt < 2:
+                # ERR_NO_BUFFER_SPACE needs the OS to recover socket
+                # buffers; back off 5s then 15s before retrying.
+                if "ERR_NO_BUFFER_SPACE" in str(e):
+                    backoff_s = 5 if _relogin_attempt == 0 else 15
+                    print(
+                        f"[ui]   ENOBUFS detected; sleeping {backoff_s}s "
+                        f"before retry to let OS recover socket buffers...",
+                        flush = True,
+                    )
+                    time.sleep(backoff_s)
+                # Replace the page if it died; otherwise next iteration's
+                # page.goto() handles the reload.
+                old_page = page
+                page = recover_or_replace_page(
+                    page,
+                    ctx,
+                    default_timeout_ms = 60_000,
+                    info = lambda m: print(f"[ui]   recovery: {m}", flush = True),
+                )
+                # A freshly created replacement page loses the pageerror/console
+                # listeners; re-attach so error tracking survives recovery.
+                if page is not old_page:
+                    page.on("pageerror", lambda e: page_errors.append(str(e)))
+                    page.on("console", _on_console)
+    if relogin_err is not None:
+        raise relogin_err
+    # Composer mount confirms the rotated session is authenticated. Kept OUTSIDE the
+    # retry: the loop breaks right after submit, so we never re-goto /login once login
+    # has set tokens -- that would hit the guest guard, redirect to /chat, and make a
+    # merely-slow composer look like a broken login.
     composer = page.locator('textarea[aria-label="Message input"]')
     composer.wait_for(state = "visible", timeout = 60_000)
     shoot("18-relogin-with-NEW2")
diff --git a/tests/studio/test_chat_response_details_ui_contract.py b/tests/studio/test_chat_response_details_ui_contract.py
index 04301a0de6..89d3000629 100644
--- a/tests/studio/test_chat_response_details_ui_contract.py
+++ b/tests/studio/test_chat_response_details_ui_contract.py
@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
 """Static contract for the chat response-details action and metadata."""
 
 from __future__ import annotations
diff --git a/tests/studio/test_stream_cancel_registration_timing.py b/tests/studio/test_stream_cancel_registration_timing.py
index 7cfe9bf32c..a833006873 100644
--- a/tests/studio/test_stream_cancel_registration_timing.py
+++ b/tests/studio/test_stream_cancel_registration_timing.py
@@ -17,7 +17,7 @@ from pathlib import Path
 
 
 SOURCE_PATH = Path(__file__).resolve().parents[2] / "studio" / "backend" / "routes" / "inference.py"
-SRC = SOURCE_PATH.read_text()
+SRC = SOURCE_PATH.read_text(encoding = "utf-8")
 _TREE = ast.parse(SRC)
 
 
@@ -166,16 +166,20 @@ def test_chat_completions_streams_avoid_starlette_task_group():
 
 
 def test_openai_passthrough_stream_avoids_starlette_task_group():
-    top = _async_function("_openai_passthrough_stream")
+    functions = [
+        _async_function("_openai_passthrough_stream"),
+        _async_function("_openai_passthrough_stream_admitted"),
+    ]
     legacy_calls = []
     same_task_calls = 0
-    for sub in ast.walk(top):
-        if not (isinstance(sub, ast.Call) and isinstance(sub.func, ast.Name)):
-            continue
-        if sub.func.id == "StreamingResponse":
-            legacy_calls.append(sub.lineno)
-        if sub.func.id == "_SameTaskStreamingResponse":
-            same_task_calls += 1
+    for fn in functions:
+        for sub in ast.walk(fn):
+            if not (isinstance(sub, ast.Call) and isinstance(sub.func, ast.Name)):
+                continue
+            if sub.func.id == "StreamingResponse":
+                legacy_calls.append(sub.lineno)
+            if sub.func.id == "_SameTaskStreamingResponse":
+                same_task_calls += 1
     assert not legacy_calls, (
         "OpenAI passthrough streams must use _SameTaskStreamingResponse, "
         "not Starlette's legacy task-group StreamingResponse. Lines: "
@@ -197,7 +201,7 @@ def test_direct_llama_server_streams_install_disconnect_watcher():
         "openai_completions",
         "_responses_stream",
         "_anthropic_passthrough_stream",
-        "_openai_passthrough_stream",
+        "_openai_passthrough_stream_admitted",
     }
     missing = [
         name
diff --git a/tests/studio/test_usage_examples_agent_detection_contract.py b/tests/studio/test_usage_examples_agent_detection_contract.py
new file mode 100644
index 0000000000..0dc16f0d30
--- /dev/null
+++ b/tests/studio/test_usage_examples_agent_detection_contract.py
@@ -0,0 +1,21 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Static contract for API usage-example agent detection scope."""
+
+from pathlib import Path
+
+REPO = Path(__file__).resolve().parents[2]
+USAGE_EXAMPLES_TSX = REPO / "studio/frontend/src/features/settings/components/usage-examples.tsx"
+
+
+def test_agent_detection_requires_desktop_scope():
+    src = USAGE_EXAMPLES_TSX.read_text(encoding = "utf-8")
+    assert 'import { isTauri } from "@/lib/api-base"' in src
+    assert "function canUseLocalAgentDetection(base: string): boolean" in src
+    helper = src[src.find("function canUseLocalAgentDetection") : src.find("const SHIKI_THEMES")]
+    assert "if (!isTauri) return false" in helper
+    assert "isLoopbackHost(normalizeHost(new URL(base).hostname))" in helper
+    assert "const localAgentDetection = canUseLocalAgentDetection(base)" in src
+    assert "if (!localAgentDetection)" in src
+    assert "}, [localAgentDetection]);" in src
diff --git a/tests/test_bad_mappings_redirect.py b/tests/test_bad_mappings_redirect.py
new file mode 100644
index 0000000000..49dab2d98b
--- /dev/null
+++ b/tests/test_bad_mappings_redirect.py
@@ -0,0 +1,47 @@
+"""Regression test for BAD_MAPPINGS redirecting oversized dynamic quants.
+
+get_model_name previously applied BAD_MAPPINGS only to the resolver's output,
+but several listed names (the `-unsloth-bnb-4bit` dynamic quants, plus any name
+the resolver doesn't map) come back as None, so their BAD_MAPPINGS entries were
+dead and the oversized model loaded. Asserting over every entry catches all of
+them. The mapper table and the resolver have no heavy imports of their own,
+so we exec the import-free mapper module and ast-extract the resolver functions
+rather than importing unsloth (which needs a GPU).
+"""
+
+import ast
+import os
+
+_MODELS = os.path.join(os.path.dirname(__file__), os.pardir, "unsloth", "models")
+
+
+def _load_get_model_name():
+    mapper_ns = {}
+    with open(os.path.join(_MODELS, "mapper.py"), encoding = "utf-8") as f:
+        exec(compile(f.read(), "mapper.py", "exec"), mapper_ns)
+
+    with open(os.path.join(_MODELS, "loader_utils.py"), encoding = "utf-8") as f:
+        tree = ast.parse(f.read())
+
+    namespace = dict(mapper_ns)
+    namespace["SUPPORTS_FOURBIT"] = True
+    namespace["_env_says_offline"] = lambda: True
+    namespace["_get_new_mapper"] = lambda: ({}, {}, {})
+
+    wanted = {"__get_model_name", "_resolve_with_mappers", "get_model_name"}
+    for node in tree.body:
+        if isinstance(node, ast.Assign) and any(
+            getattr(target, "id", None) == "BAD_MAPPINGS" for target in node.targets
+        ):
+            exec(compile(ast.Module([node], []), "<bad_mappings>", "exec"), namespace)
+        elif isinstance(node, ast.FunctionDef) and node.name in wanted:
+            exec(compile(ast.Module([node], []), node.name, "exec"), namespace)
+
+    return namespace["get_model_name"], namespace["BAD_MAPPINGS"]
+
+
+def test_bad_mappings_redirect_every_listed_name():
+    get_model_name, bad_mappings = _load_get_model_name()
+    assert bad_mappings, "BAD_MAPPINGS should not be empty"
+    for name, expected in bad_mappings.items():
+        assert get_model_name(name, load_in_4bit = True) == expected, name
diff --git a/tests/test_fast_gemv_dispatch.py b/tests/test_fast_gemv_dispatch.py
new file mode 100644
index 0000000000..7758db2cd9
--- /dev/null
+++ b/tests/test_fast_gemv_dispatch.py
@@ -0,0 +1,63 @@
+# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""`get_lora_parameters` must not treat a `weight_scale` as a quant state for a weight that is
+already dequantized to bf16 (e.g. a compressed-tensors layer at forward time). Otherwise the
+bnb fast_gemv / fast_dequantize path reads a missing `absmax` and crashes.
+"""
+
+from types import SimpleNamespace
+
+import pytest
+import torch
+
+# unsloth.kernels.utils imports bitsandbytes unconditionally, so skip the whole module up
+# front on runners without it (e.g. CPU-only) before importing unsloth, otherwise collection
+# errors instead of producing a skip. Any other import error still surfaces as a failure.
+pytest.importorskip("bitsandbytes")
+
+import unsloth  # noqa: F401  (sets UNSLOTH_IS_PRESENT before transformers)
+from unsloth.kernels.utils import get_lora_parameters_bias, _FP8_WEIGHT_DTYPES
+
+_FP8 = _FP8_WEIGHT_DTYPES[0] if _FP8_WEIGHT_DTYPES else None
+
+
+def _proj(weight, weight_scale = None):
+    proj = SimpleNamespace(weight = weight, bias = None, merged = False)
+    if weight_scale is not None:
+        proj.weight_scale = weight_scale
+    return proj
+
+
+def test_bf16_weight_scale_not_used_as_quant_state():
+    """A bf16 weight carrying a weight_scale (compressed-tensors) -> quant state must be None."""
+    proj = _proj(torch.randn(4, 4, dtype = torch.bfloat16), torch.rand(2, 2))
+    W, W_quant = get_lora_parameters_bias(proj)[:2]
+    assert W_quant is None
+
+
+def test_fp8_weight_keeps_scale():
+    """An actual fp8 weight still resolves its weight_scale as the quant state."""
+    if _FP8 is None:
+        pytest.skip("no float8 dtype in this torch build")
+    scale = torch.rand(2, 2)
+    proj = _proj(torch.randn(4, 4).to(_FP8), scale)
+    W, W_quant = get_lora_parameters_bias(proj)[:2]
+    assert W_quant is scale
+
+
+def test_plain_bf16_has_no_quant_state():
+    proj = _proj(torch.randn(4, 4, dtype = torch.bfloat16))
+    W, W_quant = get_lora_parameters_bias(proj)[:2]
+    assert W_quant is None
diff --git a/tests/test_fp8_device_context.py b/tests/test_fp8_device_context.py
new file mode 100644
index 0000000000..2eea35f4e6
--- /dev/null
+++ b/tests/test_fp8_device_context.py
@@ -0,0 +1,249 @@
+from __future__ import annotations
+
+import ast
+from contextlib import nullcontext
+from pathlib import Path
+from types import SimpleNamespace
+
+import pytest
+
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+FP8_SOURCE = REPO_ROOT / "unsloth" / "kernels" / "fp8.py"
+
+
+class _FakeDeviceModule:
+    def __init__(self, device_count: int) -> None:
+        self._device_count = device_count
+        self.device_calls = []
+
+    def device_count(self) -> int:
+        return self._device_count
+
+    def device(self, device):
+        self.device_calls.append(device)
+        return ("device-context", device)
+
+
+class _FakeTorch:
+    Tensor = object
+
+    def __init__(
+        self,
+        cuda_device_count: int,
+        xpu_device_count: int = 0,
+    ) -> None:
+        self.cuda = _FakeDeviceModule(cuda_device_count)
+        self.xpu = _FakeDeviceModule(xpu_device_count)
+
+
+class _LaunchVisitor(ast.NodeVisitor):
+    def __init__(self) -> None:
+        self.guarded_launches: set[str] = set()
+        self.unguarded_launches: set[str] = set()
+        self._inside_fp8_device_context = 0
+
+    def visit_With(self, node: ast.With) -> None:
+        enters_context = any(
+            isinstance(item.context_expr, ast.Call)
+            and isinstance(item.context_expr.func, ast.Name)
+            and item.context_expr.func.id == "_fp8_triton_device_context"
+            for item in node.items
+        )
+        if enters_context:
+            self._inside_fp8_device_context += 1
+        for statement in node.body:
+            self.visit(statement)
+        if enters_context:
+            self._inside_fp8_device_context -= 1
+
+    def visit_Call(self, node: ast.Call) -> None:
+        launch_name = self._triton_launch_name(node)
+        if launch_name is not None:
+            if self._inside_fp8_device_context:
+                self.guarded_launches.add(launch_name)
+            else:
+                self.unguarded_launches.add(launch_name)
+        self.generic_visit(node)
+
+    @staticmethod
+    def _triton_launch_name(node: ast.Call) -> str | None:
+        if isinstance(node.func, ast.Name) and node.func.id == "triton_quantize_fp8_block":
+            return node.func.id
+        if not isinstance(node.func, ast.Subscript):
+            return None
+        if not isinstance(node.func.value, ast.Name):
+            return None
+        return node.func.value.id
+
+
+def _load_device_context_helper(fake_torch: _FakeTorch):
+    source = FP8_SOURCE.read_text()
+    tree = ast.parse(source)
+    for node in tree.body:
+        if isinstance(node, ast.FunctionDef) and node.name == "_fp8_triton_device_context":
+            namespace = {"torch": fake_torch, "nullcontext": nullcontext}
+            exec(ast.get_source_segment(source, node), namespace)
+            return namespace["_fp8_triton_device_context"]
+    raise AssertionError("_fp8_triton_device_context was not found")
+
+
+def test_fp8_device_context_selects_cuda_tensor_device_on_multi_gpu() -> None:
+    fake_torch = _FakeTorch(cuda_device_count = 2)
+    helper = _load_device_context_helper(fake_torch)
+    tensor = SimpleNamespace(device = SimpleNamespace(type = "cuda"))
+
+    context = helper(tensor)
+
+    assert context == ("device-context", tensor.device)
+    assert fake_torch.cuda.device_calls == [tensor.device]
+
+
+def test_fp8_device_context_is_noop_for_single_cuda_device() -> None:
+    fake_torch = _FakeTorch(cuda_device_count = 1)
+    helper = _load_device_context_helper(fake_torch)
+    tensor = SimpleNamespace(device = SimpleNamespace(type = "cuda"))
+
+    context = helper(tensor)
+
+    assert isinstance(context, nullcontext)
+    assert fake_torch.cuda.device_calls == []
+
+
+def test_fp8_device_context_selects_xpu_tensor_device_on_multi_gpu() -> None:
+    fake_torch = _FakeTorch(cuda_device_count = 0, xpu_device_count = 2)
+    helper = _load_device_context_helper(fake_torch)
+    tensor = SimpleNamespace(device = SimpleNamespace(type = "xpu"))
+
+    context = helper(tensor)
+
+    assert context == ("device-context", tensor.device)
+    assert fake_torch.xpu.device_calls == [tensor.device]
+
+
+def test_fp8_device_context_is_noop_for_single_xpu_device() -> None:
+    fake_torch = _FakeTorch(cuda_device_count = 0, xpu_device_count = 1)
+    helper = _load_device_context_helper(fake_torch)
+    tensor = SimpleNamespace(device = SimpleNamespace(type = "xpu"))
+
+    context = helper(tensor)
+
+    assert isinstance(context, nullcontext)
+    assert fake_torch.xpu.device_calls == []
+
+
+def test_fp8_device_context_is_noop_for_non_cuda_tensor() -> None:
+    fake_torch = _FakeTorch(cuda_device_count = 8)
+    helper = _load_device_context_helper(fake_torch)
+    tensor = SimpleNamespace(device = SimpleNamespace(type = "cpu"))
+
+    context = helper(tensor)
+
+    assert isinstance(context, nullcontext)
+    assert fake_torch.cuda.device_calls == []
+
+
+def test_fp8_triton_launches_enter_tensor_device_context() -> None:
+    tree = ast.parse(FP8_SOURCE.read_text())
+    function_names = {node.name for node in ast.walk(tree) if isinstance(node, ast.FunctionDef)}
+    assert "_fp8_triton_device_context" in function_names
+
+    visitor = _LaunchVisitor()
+    visitor.visit(tree)
+
+    expected_launches = {
+        "weight_dequant_kernel",
+        "act_quant_kernel",
+        "_w8a8_block_fp8_matmul",
+        "triton_quantize_fp8_block",
+    }
+    assert expected_launches <= visitor.guarded_launches
+    assert not (expected_launches & visitor.unguarded_launches)
+
+
+def _require_two_cuda_devices():
+    torch = pytest.importorskip("torch")
+    pytest.importorskip("triton")
+
+    if not torch.cuda.is_available() or torch.cuda.device_count() < 2:
+        pytest.skip("requires at least two CUDA devices")
+    return torch
+
+
+def test_weight_dequant_block_runs_on_tensor_device_when_current_device_differs() -> None:
+    torch = _require_two_cuda_devices()
+    from unsloth.kernels.fp8 import weight_dequant_block
+
+    previous_device = torch.cuda.current_device()
+    try:
+        torch.cuda.set_device(0)
+        x = torch.arange(256 * 256, device = "cuda:1", dtype = torch.float32).reshape(256, 256)
+        scales = torch.tensor([[1.0, 2.0], [3.0, 4.0]], device = "cuda:1", dtype = torch.float32)
+
+        actual = weight_dequant_block(x, scales, block_size = 128, dtype = torch.float32)
+
+        expanded_scales = scales.repeat_interleave(128, dim = 0).repeat_interleave(128, dim = 1)
+        expected = x * expanded_scales
+
+        assert actual.device == x.device
+        assert torch.cuda.current_device() == 0
+        torch.testing.assert_close(actual, expected)
+    finally:
+        torch.cuda.set_device(previous_device)
+
+
+def test_act_quant_runs_on_tensor_device_when_current_device_differs() -> None:
+    torch = _require_two_cuda_devices()
+    if not hasattr(torch, "float8_e4m3fn"):
+        pytest.skip("requires torch.float8_e4m3fn")
+    if torch.cuda.get_device_capability(1)[0] < 9:
+        pytest.skip("requires FP8-capable CUDA hardware")
+
+    from unsloth.kernels.fp8 import act_quant
+
+    previous_device = torch.cuda.current_device()
+    try:
+        torch.cuda.set_device(0)
+        x = torch.arange(256, device = "cuda:1", dtype = torch.float32).reshape(2, 128)
+
+        y, scales = act_quant(x, block_size = 128)
+
+        assert y.device == x.device
+        assert scales.device == x.device
+        assert torch.cuda.current_device() == 0
+    finally:
+        torch.cuda.set_device(previous_device)
+
+
+def test_w8a8_block_fp8_matmul_triton_runs_on_tensor_device_when_current_device_differs() -> None:
+    torch = _require_two_cuda_devices()
+    if not hasattr(torch, "float8_e4m3fn"):
+        pytest.skip("requires torch.float8_e4m3fn")
+    if torch.cuda.get_device_capability(1)[0] < 9:
+        pytest.skip("requires FP8-capable CUDA hardware")
+
+    from unsloth.kernels.fp8 import w8a8_block_fp8_matmul_triton
+
+    previous_device = torch.cuda.current_device()
+    try:
+        torch.cuda.set_device(0)
+        A = torch.ones((128, 128), device = "cuda:1", dtype = torch.float32).to(torch.float8_e4m3fn)
+        B = torch.ones((128, 128), device = "cuda:1", dtype = torch.float32).to(torch.float8_e4m3fn)
+        As = torch.ones((128, 1), device = "cuda:1", dtype = torch.float32)
+        Bs = torch.ones((1, 1), device = "cuda:1", dtype = torch.float32)
+
+        actual = w8a8_block_fp8_matmul_triton(
+            A,
+            B,
+            As,
+            Bs,
+            block_size = [128, 128],
+            output_dtype = torch.float32,
+        )
+
+        expected = torch.full((128, 128), 128.0, device = "cuda:1", dtype = torch.float32)
+        assert actual.device == A.device
+        assert torch.cuda.current_device() == 0
+        torch.testing.assert_close(actual, expected)
+    finally:
+        torch.cuda.set_device(previous_device)
diff --git a/tests/test_fp8_restore_dropped_scale.py b/tests/test_fp8_restore_dropped_scale.py
new file mode 100644
index 0000000000..a85b24daff
--- /dev/null
+++ b/tests/test_fp8_restore_dropped_scale.py
@@ -0,0 +1,358 @@
+# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Restoring dropped block-fp8 `weight_scale_inv` tensors on load (#6200).
+
+Some block-scale fp8 checkpoints leave a Linear (e.g. `mlp.gate_proj`) unconverted, so its raw
+quantized values land in a plain bf16 weight and its `weight_scale_inv` is dropped, producing a
+garbage un-scaled weight. `_restore_dropped_fp8_scales` dequantizes such orphaned weights in place
+using the scale from the checkpoint. Runs offline on CPU with synthetic checkpoints.
+"""
+
+import json
+import os
+import tempfile
+from types import SimpleNamespace
+
+import torch
+from torch import nn
+from safetensors.torch import save_file
+
+# Import unsloth first to set UNSLOTH_IS_PRESENT env var.
+import unsloth
+from unsloth.models.loader_utils import _restore_dropped_fp8_scales, _FP8_DTYPES
+
+
+_SHARD = "model-00001-of-00001.safetensors"
+_FP8 = _FP8_DTYPES[0] if _FP8_DTYPES else None
+
+
+def _write_checkpoint(
+    path,
+    tensors,
+    filename = _SHARD,
+    include_index = True,
+):
+    save_file(tensors, os.path.join(path, filename))
+    if include_index:
+        weight_map = {name: filename for name in tensors}
+        with open(os.path.join(path, "model.safetensors.index.json"), "w") as f:
+            json.dump({"weight_map": weight_map}, f)
+
+
+def _fp8_config(block = (2, 2)):
+    return SimpleNamespace(
+        quantization_config = {
+            "quant_method": "fp8",
+            "weight_block_size": list(block),
+        }
+    )
+
+
+def _fp8_anchor():
+    """A module carrying a real fp8 weight, so the model looks like a genuine fp8 load."""
+    m = nn.Linear(2, 2, bias = False)
+    m.weight = nn.Parameter(torch.randn(2, 2).to(_FP8), requires_grad = False)
+    return m
+
+
+def _bf16_linear(out_f, in_f, raw):
+    m = nn.Linear(in_f, out_f, bias = False).to(torch.bfloat16)
+    with torch.no_grad():
+        m.weight.copy_(raw)
+    return m
+
+
+def _expand(scale, block, shape):
+    bs0, bs1 = block
+    expanded = scale.repeat_interleave(bs0, dim = 0).repeat_interleave(bs1, dim = 1)
+    return expanded[: shape[0], : shape[1]]
+
+
+def test_restore_dequantizes_orphaned_scale():
+    """A plain bf16 weight whose scale was dropped is dequantized in place."""
+    if _FP8 is None:
+        return
+    torch.manual_seed(0)
+    raw = torch.randn(4, 4, dtype = torch.bfloat16)
+    scale = torch.rand(2, 2, dtype = torch.float32) + 0.1
+
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.anchor = _fp8_anchor()
+    model.layer = _bf16_linear(4, 4, raw)
+
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(
+            d,
+            {
+                "layer.weight": raw.to(torch.float32),
+                "layer.weight_scale_inv": scale,
+            },
+        )
+        restored, skipped = _restore_dropped_fp8_scales(model, d, local_files_only = True)
+
+    assert restored == 1
+    expected = (raw.to(torch.float32) * _expand(scale, (2, 2), (4, 4))).to(torch.bfloat16)
+    assert torch.equal(model.layer.weight.data, expected)
+
+
+def test_skips_already_fp8_weight():
+    """A correctly converted fp8 weight is skipped, never double-scaled."""
+    if _FP8 is None:
+        return
+    weight = torch.randn(4, 4).to(_FP8)
+    before = weight.clone()
+
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.layer = nn.Linear(4, 4, bias = False)
+    model.layer.weight = nn.Parameter(weight, requires_grad = False)
+
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(d, {"layer.weight_scale_inv": torch.rand(2, 2)})
+        restored, skipped = _restore_dropped_fp8_scales(model, d, local_files_only = True)
+
+    assert restored == 0 and skipped == 1
+    assert torch.equal(model.layer.weight.data.float(), before.float())
+
+
+def test_skips_offloaded_meta_weight():
+    """A disk-offloaded layer (weight on the meta device) is skipped without error or restore."""
+    if _FP8 is None:
+        return
+    raw = torch.randn(4, 4, dtype = torch.bfloat16)
+    scale = torch.rand(2, 2, dtype = torch.float32) + 0.1
+
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.anchor = _fp8_anchor()
+    model.layer = nn.Linear(4, 4, bias = False)
+    # Simulate an offloaded weight living on the meta device.
+    model.layer.weight = nn.Parameter(
+        torch.empty(4, 4, dtype = torch.bfloat16, device = "meta"), requires_grad = False
+    )
+
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(
+            d,
+            {
+                "layer.weight": raw.to(torch.float32),
+                "layer.weight_scale_inv": scale,
+            },
+        )
+        restored, skipped = _restore_dropped_fp8_scales(model, d, local_files_only = True)
+
+    assert restored == 0
+    assert model.layer.weight.device.type == "meta"
+
+
+def test_noop_when_fully_dequantized():
+    """If the model has no fp8 weights at all (e.g. load_in_16bit dequantize), do not rescale."""
+    raw = torch.randn(4, 4, dtype = torch.bfloat16)
+    scale = torch.rand(2, 2, dtype = torch.float32) + 0.1
+
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.layer = _bf16_linear(4, 4, raw)  # no fp8 anchor -> looks dequantized
+
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(d, {"layer.weight_scale_inv": scale})
+        restored, skipped = _restore_dropped_fp8_scales(model, d, local_files_only = True)
+
+    assert (restored, skipped) == (0, 0)
+    assert torch.equal(model.layer.weight.data, raw)  # untouched
+
+
+def test_non_block_divisible_shape():
+    """Block scale is expanded then sliced to a non-divisible weight shape."""
+    if _FP8 is None:
+        return
+    raw = torch.randn(3, 4, dtype = torch.bfloat16)
+    scale = torch.rand(2, 2, dtype = torch.float32) + 0.1
+
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.anchor = _fp8_anchor()
+    model.layer = _bf16_linear(3, 4, raw)  # weight shape [3, 4]
+
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(d, {"layer.weight_scale_inv": scale})
+        restored, skipped = _restore_dropped_fp8_scales(model, d, local_files_only = True)
+
+    assert restored == 1
+    expected = (raw.to(torch.float32) * _expand(scale, (2, 2), (3, 4))).to(torch.bfloat16)
+    assert torch.equal(model.layer.weight.data, expected)
+
+
+def test_transposed_scale_layout():
+    """A scale stored in the transposed block grid is transposed before use."""
+    if _FP8 is None:
+        return
+    raw = torch.randn(4, 2, dtype = torch.bfloat16)  # weight [4, 2] -> grid (2, 1)
+    scale_correct = torch.rand(2, 1, dtype = torch.float32) + 0.1
+    scale_stored = scale_correct.t().contiguous()  # stored transposed as (1, 2)
+
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.anchor = _fp8_anchor()
+    model.layer = _bf16_linear(4, 2, raw)
+
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(d, {"layer.weight_scale_inv": scale_stored})
+        restored, _ = _restore_dropped_fp8_scales(model, d, local_files_only = True)
+
+    assert restored == 1
+    expected = (raw.to(torch.float32) * _expand(scale_correct, (2, 2), (4, 2))).to(torch.bfloat16)
+    assert torch.equal(model.layer.weight.data, expected)
+
+
+def test_single_file_checkpoint_without_index():
+    """Unsharded model.safetensors (no index) is still scanned for dropped scales."""
+    if _FP8 is None:
+        return
+    raw = torch.randn(4, 4, dtype = torch.bfloat16)
+    scale = torch.rand(2, 2, dtype = torch.float32) + 0.1
+
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.anchor = _fp8_anchor()
+    model.layer = _bf16_linear(4, 4, raw)
+
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(
+            d, {"layer.weight_scale_inv": scale}, filename = "model.safetensors", include_index = False
+        )
+        restored, _ = _restore_dropped_fp8_scales(model, d, local_files_only = True)
+
+    assert restored == 1
+    expected = (raw.to(torch.float32) * _expand(scale, (2, 2), (4, 4))).to(torch.bfloat16)
+    assert torch.equal(model.layer.weight.data, expected)
+
+
+def test_scalar_block_size_config():
+    """A scalar weight_block_size (not a list) is handled without error."""
+    if _FP8 is None:
+        return
+    raw = torch.randn(4, 4, dtype = torch.bfloat16)
+    scale = torch.rand(2, 2, dtype = torch.float32) + 0.1
+
+    model = nn.Module()
+    model.config = SimpleNamespace(
+        quantization_config = {"quant_method": "fp8", "weight_block_size": 2}
+    )
+    model.anchor = _fp8_anchor()
+    model.layer = _bf16_linear(4, 4, raw)
+
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(d, {"layer.weight_scale_inv": scale})
+        restored, _ = _restore_dropped_fp8_scales(model, d, local_files_only = True)
+
+    assert restored == 1
+
+
+def test_text_only_prefix_mapping():
+    """Checkpoint keys with a language_model prefix match the stripped text-only module names."""
+    if _FP8 is None:
+        return
+    raw = torch.randn(2, 2, dtype = torch.bfloat16)
+    scale = torch.rand(1, 1, dtype = torch.float32) + 0.1
+
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.anchor = _fp8_anchor()
+    model.model = nn.Module()
+    model.model.gate_proj = _bf16_linear(2, 2, raw)  # module lacks the language_model prefix
+
+    with tempfile.TemporaryDirectory() as d:
+        # checkpoint key carries the language_model wrapper the text-only load stripped
+        _write_checkpoint(d, {"model.language_model.gate_proj.weight_scale_inv": scale})
+        restored, _ = _restore_dropped_fp8_scales(model, d, local_files_only = True)
+
+    assert restored == 1
+    expected = (raw.to(torch.float32) * _expand(scale, (2, 2), (2, 2))).to(torch.bfloat16)
+    assert torch.equal(model.model.gate_proj.weight.data, expected)
+
+
+def test_skips_variant_load():
+    """A variant load (variant="fp8") is skipped to avoid applying default-checkpoint scales."""
+    if _FP8 is None:
+        return
+    raw = torch.randn(4, 4, dtype = torch.bfloat16)
+    scale = torch.rand(2, 2, dtype = torch.float32) + 0.1
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.anchor = _fp8_anchor()
+    model.layer = _bf16_linear(4, 4, raw)
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(d, {"layer.weight_scale_inv": scale})
+        result = _restore_dropped_fp8_scales(model, d, local_files_only = True, variant = "fp8")
+    assert result == (0, 0)
+    assert torch.equal(model.layer.weight.data, raw)  # untouched
+
+
+def test_vlm_language_model_model_alias():
+    """A checkpoint key language_model.model.* matches a model.language_model.* module."""
+    if _FP8 is None:
+        return
+    raw = torch.randn(2, 2, dtype = torch.bfloat16)
+    scale = torch.rand(1, 1, dtype = torch.float32) + 0.1
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.anchor = _fp8_anchor()
+    model.model = nn.Module()
+    model.model.language_model = nn.Module()
+    model.model.language_model.gate_proj = _bf16_linear(
+        2, 2, raw
+    )  # -> model.language_model.gate_proj
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(d, {"language_model.model.gate_proj.weight_scale_inv": scale})
+        restored, _ = _restore_dropped_fp8_scales(model, d, local_files_only = True)
+    assert restored == 1
+    expected = (raw.to(torch.float32) * _expand(scale, (2, 2), (2, 2))).to(torch.bfloat16)
+    assert torch.equal(model.model.language_model.gate_proj.weight.data, expected)
+
+
+def test_noop_without_scale_keys():
+    if _FP8 is None:
+        return
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.anchor = _fp8_anchor()
+    model.layer = _bf16_linear(4, 4, torch.randn(4, 4, dtype = torch.bfloat16))
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(d, {"layer.weight": torch.randn(4, 4)})
+        assert _restore_dropped_fp8_scales(model, d, local_files_only = True) == (0, 0)
+
+
+def test_noop_without_index_or_single_file():
+    if _FP8 is None:
+        return
+    model = nn.Module()
+    model.config = _fp8_config((2, 2))
+    model.anchor = _fp8_anchor()
+    model.layer = _bf16_linear(4, 4, torch.randn(4, 4, dtype = torch.bfloat16))
+    with tempfile.TemporaryDirectory() as d:
+        assert _restore_dropped_fp8_scales(model, d, local_files_only = True) == (0, 0)
+
+
+def test_noop_when_not_block_fp8():
+    """A non-fp8 (or non-block) quantization config is ignored."""
+    scale = torch.rand(2, 2)
+    model = nn.Module()
+    model.config = SimpleNamespace(quantization_config = {"quant_method": "compressed-tensors"})
+    model.layer = nn.Linear(4, 4, bias = False)
+    with tempfile.TemporaryDirectory() as d:
+        _write_checkpoint(d, {"layer.weight_scale_inv": scale})
+        assert _restore_dropped_fp8_scales(model, d, local_files_only = True) == (0, 0)
diff --git a/tests/test_gemma_2b_mapper_key.py b/tests/test_gemma_2b_mapper_key.py
new file mode 100644
index 0000000000..31edacfce4
--- /dev/null
+++ b/tests/test_gemma_2b_mapper_key.py
@@ -0,0 +1,46 @@
+"""Regression test for the duplicate ``unsloth/gemma-2b-bnb-4bit`` key in
+``unsloth/models/mapper.py``.
+
+The 4bit instruction-tuned Gemma 2B entry was accidentally keyed with the base
+model's repo name, so ``__INT_TO_FLOAT_MAPPER`` held two identical
+``unsloth/gemma-2b-bnb-4bit`` keys. Python keeps only the last value for a
+duplicate literal key, so the base 4bit repo resolved to the *instruct* model,
+the base model lost its reverse (4x-faster) mapping, and
+``unsloth/gemma-2b-it-bnb-4bit`` was never registered at all.
+
+``mapper.py`` has no imports, so we exec it directly and inspect the built
+mappers without importing ``unsloth`` (which requires a GPU).
+"""
+
+import os
+
+MAPPER_PATH = os.path.join(os.path.dirname(__file__), os.pardir, "unsloth", "models", "mapper.py")
+
+
+def _load_mappers():
+    with open(MAPPER_PATH) as f:
+        source = f.read()
+    namespace = {}
+    exec(compile(source, MAPPER_PATH, "exec"), namespace)
+    return namespace
+
+
+def test_gemma_2b_base_and_instruct_4bit_are_distinct():
+    namespace = _load_mappers()
+    int_to_float = namespace["INT_TO_FLOAT_MAPPER"]
+    float_to_int = namespace["FLOAT_TO_INT_MAPPER"]
+
+    # The base 4bit repo must resolve to the base model, not the instruct one.
+    assert int_to_float["unsloth/gemma-2b-bnb-4bit"] == "unsloth/gemma-2b"
+
+    # The instruct 4bit repo must be registered and resolve to the instruct model.
+    assert "unsloth/gemma-2b-it-bnb-4bit" in int_to_float
+    assert int_to_float["unsloth/gemma-2b-it-bnb-4bit"] == "unsloth/gemma-2b-it"
+
+    # The base model must reverse-map back to the base 4bit repo.
+    assert float_to_int["unsloth/gemma-2b"] == "unsloth/gemma-2b-bnb-4bit"
+    assert float_to_int["google/gemma-2b"] == "unsloth/gemma-2b-bnb-4bit"
+
+    # The instruct model must reverse-map to the instruct 4bit repo.
+    assert float_to_int["unsloth/gemma-2b-it"] == "unsloth/gemma-2b-it-bnb-4bit"
+    assert float_to_int["google/gemma-2b-it"] == "unsloth/gemma-2b-it-bnb-4bit"
diff --git a/tests/utils/test_rope_scaling_drift.py b/tests/utils/test_rope_scaling_drift.py
index 7a738e236c..b2ec1e5a20 100644
--- a/tests/utils/test_rope_scaling_drift.py
+++ b/tests/utils/test_rope_scaling_drift.py
@@ -257,6 +257,39 @@ def test_recompute_helper_scales_on_cpu():
     ), "_unsloth_recompute_inv_freq must return vanilla inv_freq when unscaled."
 
 
+def test_extended_rope_scaling_keeps_llama3_and_carries_theta():
+    # Long-context extension keeps native llama3, but falls back to linear for every other
+    # type (the patched attention constructor only rebuilds linear/llama3/longrope), and the
+    # linear dict carries rope_theta so transformers v5 does not fall back to base 10000.
+    from types import SimpleNamespace
+
+    from unsloth.models.llama import _extended_rope_scaling
+
+    # llama3 model: keep native scaling, do not synthesize linear.
+    scaling, native = _extended_rope_scaling(_make_config(LLAMA3_ROPE_SCALING), 2.0)
+    assert (
+        scaling is None and native == "llama3"
+    ), "must keep native llama3 scaling instead of overwriting it with linear."
+
+    # yarn is not rebuildable by the patcher -> keep the safe linear fallback, not native.
+    yarn = SimpleNamespace(rope_scaling = {"rope_type": "yarn", "factor": 2.0}, rope_theta = 500000.0)
+    scaling, _ = _extended_rope_scaling(yarn, 2.0)
+    assert scaling == {
+        "type": "linear",
+        "factor": 2.0,
+        "rope_theta": 500000.0,
+    }, f"yarn must fall back to linear (patcher cannot rebuild it), got {scaling}."
+
+    # plain RoPE with theta only under v5 rope_parameters: linear must carry rope_theta.
+    v5 = SimpleNamespace(rope_parameters = {"rope_type": "default", "rope_theta": 1000000.0})
+    scaling, _ = _extended_rope_scaling(v5, 2.0)
+    assert scaling == {
+        "type": "linear",
+        "factor": 2.0,
+        "rope_theta": 1000000.0,
+    }, f"linear override dropped rope_theta on v5 (got {scaling}); base would fall back to 10000."
+
+
 def test_extended_rotary_reads_config_factor():
     # LlamaExtendedRotaryEmbedding must honor the config factor, not hardcode 8
     # (Llama-3.2 uses 32); otherwise the subclass path re-drops scaling (#2405).
diff --git a/tests/version_compat/test_sentence_transformers_pinned_symbols.py b/tests/version_compat/test_sentence_transformers_pinned_symbols.py
index c0c35b9d5d..d7f9a54811 100644
--- a/tests/version_compat/test_sentence_transformers_pinned_symbols.py
+++ b/tests/version_compat/test_sentence_transformers_pinned_symbols.py
@@ -19,6 +19,8 @@ ST_TAGS = [
     "v5.2.3",
     "v5.3.0",
     "v5.4.1",
+    "v5.5.1",
+    "v5.6.0",
     "master",
 ]
 
@@ -120,6 +122,42 @@ def test_st_transformer_base_class_either_path(tag: str):
     )
 
 
+# Transformer.load classmethod: unsloth builds saved-ST modules through it (#6881).
+@pytest.mark.parametrize("tag", ST_TAGS)
+def test_st_transformer_load_accepts_unsloth_kwargs(tag: str):
+    """unsloth builds saved ST models via Transformer.load(...) so the saved
+    modality_config is honored (#6881). If .load stops accepting the hub kwargs it
+    passes (and has no **kwargs), update the fix before it silently regresses. Not
+    locating .load is a SKIP (may be inherited); the live test guards the install."""
+    candidates = [
+        "sentence_transformers/models/Transformer.py",
+        "sentence_transformers/models/transformer.py",
+        "sentence_transformers/base/modules/transformer.py",
+        "sentence_transformers/base/modules/module.py",
+    ]
+    for p in candidates:
+        src = fetch_text("UKPLab/sentence-transformers", tag, p)
+        if src is None or not has_def(src, "load", "func"):
+            continue
+        m = re.search(r"def\s+load\s*\((.*?)\)\s*(?:->[^:]*)?:", src, re.S)
+        if m is None:
+            continue
+        sig = m.group(1)
+        accepts_var_kw = "**" in sig
+        missing = [
+            kw
+            for kw in ("token", "cache_folder", "revision", "trust_remote_code")
+            if not (accepts_var_kw or re.search(rf"\b{re.escape(kw)}\b", sig))
+        ]
+        assert not missing, (
+            f"{tag}: Transformer.load in {p} no longer accepts {missing} and has no "
+            f"**kwargs; update unsloth.models.sentence_transformer._create_transformer_module "
+            f"(#6881) before it silently falls back to Transformer(...)."
+        )
+        return
+    pytest.skip(f"{tag}: Transformer.load not locatable in {candidates} (may be inherited)")
+
+
 # sentence_transformers.util: import_from_string + load_dir_path helpers unsloth calls.
 @pytest.mark.parametrize("tag", ST_TAGS)
 def test_st_util_helpers(tag: str):
diff --git a/tests/version_compat/test_trl_fake_train_cpu.py b/tests/version_compat/test_trl_fake_train_cpu.py
new file mode 100644
index 0000000000..4dae696282
--- /dev/null
+++ b/tests/version_compat/test_trl_fake_train_cpu.py
@@ -0,0 +1,285 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team.
+"""Fake CPU training runs for the Unsloth-patched SFT / GRPO / DPO trainers.
+
+The patch-run canary (test_trl_grpo_fake_run.py) only compiles + inspects the
+generated trainer source. This goes one layer deeper: it actually runs
+`trainer.train()` for a couple of steps on a CPU-only runner, under the CUDA
+spoof, wrapping a plain (tiny, random-weight) HF model in the Unsloth-patched
+trainer. That exercises the real train() loop at runtime -- data collation,
+generation (GRPO), the injected `_get_per_token_logps_and_entropies`, loss,
+backward, optimizer -- so a TRL or transformers change that breaks the loop
+(not just the source structure) surfaces here. No GPU, no meaningful numerics.
+
+What it does NOT cover: Unsloth's Triton/GPU-optimized model kernels (the
+FastLanguageModel fast path) cannot run on CPU, so this validates the
+trainer-transform + orchestration layer with a standard forward, not the
+optimized kernels.
+"""
+
+from __future__ import annotations
+
+import os
+
+# CPU-only: no torch.compile / dynamo (it reaches into the CUDA accelerator), no
+# Unsloth kernel compile, no mixed precision. Must be set before torch/unsloth.
+os.environ.setdefault("UNSLOTH_COMPILE_DISABLE", "1")
+os.environ.setdefault("TORCHDYNAMO_DISABLE", "1")
+os.environ.setdefault("TORCH_COMPILE_DISABLE", "1")
+os.environ.setdefault("ACCELERATE_MIXED_PRECISION", "no")
+
+import importlib
+import importlib.util
+import sys
+from pathlib import Path
+
+import pytest
+
+
+# torch is needed for everything below (daily-fresh-fetch collects this dir with
+# only pytest installed); skip the whole module cleanly when it is absent.
+if importlib.util.find_spec("torch") is None:
+    pytest.skip(
+        "torch not installed; fake CPU train needs the real runtime", allow_module_level = True
+    )
+
+# Apply the CUDA spoof before any unsloth-touching import.
+_SPOOF_DIR = Path(__file__).resolve().parents[1]
+sys.path.insert(0, str(_SPOOF_DIR))
+import _zoo_aggressive_cuda_spoof as _spoof  # noqa: E402
+
+_spoof.apply()
+
+import torch  # noqa: E402
+
+
+# The generated GRPO trainer hard-decorates hot functions with @torch.compile,
+# which dynamo processes even under the disable env vars, reaching into
+# torch.accelerator (real CUDA) on a GPU-less box. Make torch.compile an eager
+# passthrough before unsloth generates/imports the trainer -- same logic, no
+# dynamo. (An eager CPU run is exactly what we want here.)
+def _eager_compile(
+    model = None,
+    *args,
+    **kwargs,
+):
+    if callable(model):
+        return model
+    return lambda fn: fn
+
+
+torch.compile = _eager_compile
+
+# Belt-and-suspenders: if any @torch.compile still routes through dynamo, let it
+# fall back to eager instead of crashing, and stop its stream-capture probe from
+# reaching torch.accelerator -> real CUDA on a GPU-less box.
+try:
+    import torch._dynamo  # noqa: E402
+    torch._dynamo.config.suppress_errors = True
+except Exception:
+    pass
+if hasattr(torch, "accelerator"):
+    torch.accelerator.is_available = lambda *a, **k: False
+
+
+# Redirect any `device="cuda"` tensor allocation / `.to("cuda")` / `.cuda()` to
+# CPU. The aggressive spoof deliberately keeps real allocators, but a fake CPU
+# train needs cuda-targeted ops (e.g. inductor's init_gpu_context does
+# `torch.empty(1, device="cuda")`) to land on CPU instead of erroring.
+def _is_cuda_dev(d):
+    try:
+        return d is not None and torch.device(d).type == "cuda"
+    except Exception:
+        return False
+
+
+for _name in (
+    "empty",
+    "zeros",
+    "ones",
+    "full",
+    "tensor",
+    "arange",
+    "randn",
+    "rand",
+    "randint",
+    "empty_like",
+    "zeros_like",
+    "ones_like",
+):
+    _orig = getattr(torch, _name, None)
+    if _orig is None:
+        continue
+
+    def _redir(
+        *args,
+        _orig = _orig,
+        **kwargs,
+    ):
+        if _is_cuda_dev(kwargs.get("device")):
+            kwargs["device"] = "cpu"
+        return _orig(*args, **kwargs)
+
+    setattr(torch, _name, _redir)
+
+_orig_to = torch.Tensor.to
+
+
+def _to_cpu(self, *args, **kwargs):
+    args = tuple("cpu" if _is_cuda_dev(a) else a for a in args)
+    if _is_cuda_dev(kwargs.get("device")):
+        kwargs["device"] = "cpu"
+    return _orig_to(self, *args, **kwargs)
+
+
+torch.Tensor.to = _to_cpu
+torch.Tensor.cuda = lambda self, *a, **k: self
+
+# Extra CUDA stubs the aggressive spoof lacks, needed to walk a real train():
+# Adam's _cuda_graph_capture_health_check() probes stream capture.
+torch.cuda.is_current_stream_capturing = lambda *a, **k: False
+try:
+    import torch.cuda.graphs as _cg  # noqa: E402
+    _cg._cuda_isCurrentStreamCapturing = lambda *a, **k: False
+except Exception:
+    pass
+
+# A broken libmlx.so in the shared site-packages crashes transformers' Mac-only
+# is_mlx_array probe on Linux; disable it.
+try:
+    import transformers.utils.generic as _g  # noqa: E402
+    _g._is_mlx_available = False
+except Exception:
+    pass
+
+
+# Dense (non-MoE) tiny model on purpose: MoE models route through Unsloth's
+# grouped_gemm Triton kernel, which is CUDA-only and cannot run on a CPU runner.
+_MODEL = "hf-internal-testing/tiny-random-LlamaForCausalLM"
+
+
+def _load_plain():
+    """Tiny plain HF model + tokenizer on CPU. Skips (not fails) if the model
+    cannot be fetched -- that is a network/hub issue, not an unsloth regression."""
+    from transformers import AutoModelForCausalLM, AutoTokenizer
+
+    try:
+        tok = AutoTokenizer.from_pretrained(_MODEL)
+        model = AutoModelForCausalLM.from_pretrained(_MODEL, dtype = torch.float32)
+    except OSError as e:  # hub unreachable / model missing
+        pytest.skip(f"could not fetch {_MODEL} (network/hub): {str(e)[:150]}")
+    if tok.pad_token is None:
+        tok.pad_token = tok.eos_token
+    # Unsloth's GRPO path calls model.for_training()/for_inference() (added by
+    # FastLanguageModel). A plain HF model lacks them; supply minimal train/eval
+    # equivalents so the loop proceeds without the optimized wrapper.
+    if not hasattr(model, "for_training"):
+        model.for_training = lambda *a, **k: model.train()
+    if not hasattr(model, "for_inference"):
+        model.for_inference = lambda *a, **k: model.eval()
+    return model.to("cpu"), tok
+
+
+@pytest.fixture(autouse = True)
+def _require_stack():
+    global torch  # the `import torch._dynamo` below would otherwise shadow it as local
+    if importlib.util.find_spec("unsloth") is None or importlib.util.find_spec("trl") is None:
+        pytest.skip("unsloth or trl not installed")
+    # A real import failure is a regression we want to surface, so do not guard it.
+    import unsloth  # noqa: F401  -- patches TRL trainers to the Unsloth variants
+
+    # `import unsloth` reinstalls the real torch.compile (overwriting the eager
+    # passthrough set at module load), so the GRPO hot path (chunked_selective_
+    # log_softmax) would really compile -- and inductor picks the spoofed CUDA
+    # device, crashing on device props (`gcnArchName`). Re-apply the eager
+    # passthrough and flip dynamo's call-time kill switch so every @torch.compile
+    # runs eager regardless of when it was decorated. CPU eager is what we want.
+    torch.compile = _eager_compile
+    try:
+        import torch._dynamo  # noqa: E402
+        torch._dynamo.config.disable = True
+    except Exception:
+        pass
+
+
+def test_sft_trains_on_cpu(tmp_path):
+    from datasets import Dataset
+    from trl import SFTConfig, SFTTrainer
+
+    assert SFTTrainer.__name__ == "UnslothSFTTrainer", "SFT patch did not apply"
+    model, tok = _load_plain()
+    ds = Dataset.from_list([{"text": "The quick brown fox jumps over the lazy dog."}] * 8)
+    cfg = SFTConfig(
+        output_dir = str(tmp_path / "ci_sft"),
+        per_device_train_batch_size = 2,
+        max_steps = 2,
+        logging_steps = 1,
+        report_to = "none",
+        save_strategy = "no",
+        use_cpu = True,
+        max_length = None,
+        padding_free = False,
+        dataset_text_field = "text",
+        fp16 = False,
+        bf16 = False,
+        optim = "adamw_torch",
+    )
+    SFTTrainer(model = model, processing_class = tok, args = cfg, train_dataset = ds).train()
+
+
+def test_grpo_trains_on_cpu(tmp_path):
+    from datasets import Dataset
+    from trl import GRPOConfig, GRPOTrainer
+
+    assert GRPOTrainer.__name__ == "UnslothGRPOTrainer", "GRPO patch did not apply"
+    model, tok = _load_plain()
+    ds = Dataset.from_list([{"prompt": "hi there"}] * 4)
+    cfg = GRPOConfig(
+        output_dir = str(tmp_path / "ci_grpo"),
+        per_device_train_batch_size = 2,
+        num_generations = 2,
+        max_steps = 2,
+        max_completion_length = 8,
+        logging_steps = 1,
+        report_to = "none",
+        temperature = 1.0,
+        beta = 0.0,
+        save_strategy = "no",
+        use_cpu = True,
+        use_vllm = False,
+        fp16 = False,
+        bf16 = False,
+        optim = "adamw_torch",
+    )
+    GRPOTrainer(
+        model = model,
+        processing_class = tok,
+        reward_funcs = [lambda completions, **k: [float(len(c)) for c in completions]],
+        args = cfg,
+        train_dataset = ds,
+    ).train()
+
+
+def test_dpo_trains_on_cpu(tmp_path):
+    from datasets import Dataset
+    from trl import DPOConfig, DPOTrainer
+
+    assert DPOTrainer.__name__ == "UnslothDPOTrainer", "DPO patch did not apply"
+    model, tok = _load_plain()
+    ds = Dataset.from_list(
+        [{"prompt": "Hi", "chosen": " hello friend", "rejected": " go away"}] * 8
+    )
+    cfg = DPOConfig(
+        output_dir = str(tmp_path / "ci_dpo"),
+        per_device_train_batch_size = 2,
+        max_steps = 2,
+        logging_steps = 1,
+        report_to = "none",
+        save_strategy = "no",
+        use_cpu = True,
+        beta = 0.1,
+        fp16 = False,
+        bf16 = False,
+        optim = "adamw_torch",
+    )
+    DPOTrainer(model = model, processing_class = tok, args = cfg, train_dataset = ds).train()
diff --git a/tests/version_compat/test_trl_grpo_fake_run.py b/tests/version_compat/test_trl_grpo_fake_run.py
new file mode 100644
index 0000000000..c79c9955da
--- /dev/null
+++ b/tests/version_compat/test_trl_grpo_fake_run.py
@@ -0,0 +1,246 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team.
+"""Fake-CUDA GRPO patch run against the *installed* TRL (CPU-only, no training).
+
+The static symbol/source-string canaries (test_trl_grpo_pinned_symbols.py)
+grep raw TRL source; they never execute unsloth's transforms. This test drives
+the real pipeline: under the aggressive CUDA spoof it imports unsloth and calls
+`_patch_trl_rl_trainers_impl`, which reads the installed GRPOTrainer via
+inspect.getsource, applies every rl.py/rl_replacements.py rewrite, and compiles
+the result into an UnslothGRPOTrainer. A structural TRL change that slips past
+the greps (e.g. TRL 1.7.0's 2->3-tuple return arity, or a restructured PEFT
+ref-adapter block) surfaces here as a transform error, a broken generated
+source, or a violated contract -- with no GPU and no training run.
+
+Meant to run in CI against `trl==latest` and `trl @ main` (see
+version-compat-ci.yml). The tests/conftest.py harness pre-loads device_type
+with DEVICE_COUNT=0 so unsloth's kernel init takes the CPU-safe path.
+"""
+
+from __future__ import annotations
+
+import ast
+import importlib
+import importlib.machinery
+import importlib.util
+import inspect
+import sys
+import types
+from pathlib import Path
+
+import pytest
+
+
+# daily-fresh-fetch collects tests/version_compat/ with only pytest installed;
+# the spoof and the rest of this module need the real torch runtime. Skip the
+# whole module cleanly when torch is absent rather than crashing collection.
+if importlib.util.find_spec("torch") is None:
+    pytest.skip("torch not installed; fake-run needs the real runtime", allow_module_level = True)
+
+# Apply the spoof BEFORE any unsloth-touching import (mirrors
+# tests/vllm_compat/test_extended_module_imports.py).
+_SPOOF_DIR = Path(__file__).resolve().parents[1]
+sys.path.insert(0, str(_SPOOF_DIR))
+import _zoo_aggressive_cuda_spoof as _spoof  # noqa: E402
+
+_spoof.apply()
+
+
+def _stub_module(name: str, attrs: dict | None = None) -> None:
+    if name in sys.modules:
+        return
+    m = types.ModuleType(name)
+    m.__spec__ = importlib.machinery.ModuleSpec(name = name, loader = None, origin = "<test stub>")
+    for k, v in (attrs or {}).items():
+        setattr(m, k, v)
+    sys.modules[name] = m
+
+
+_stub_module("torchcodec")
+
+
+def _trl_version():
+    import trl
+    from packaging.version import Version
+    return Version(trl.__version__.split("+")[0])
+
+
+def _patch_grpo_and_get_source() -> str:
+    """Run the GRPO patcher against the installed TRL and return the generated
+    UnslothGRPOTrainer source. Calls the impl (not the try/except wrapper) so a
+    transform/compile regression surfaces as a hard error instead of a silent
+    no-op."""
+    import trl.trainer.grpo_trainer as _g
+
+    from unsloth.models import rl as _rl
+
+    _rl._patch_trl_rl_trainers_impl("grpo_trainer")
+    patched = _g.GRPOTrainer
+    assert patched.__name__ == "UnslothGRPOTrainer", (
+        f"GRPO patch silently no-oped: trl.trainer.grpo_trainer.GRPOTrainer is "
+        f"{patched.__name__!r}, expected 'UnslothGRPOTrainer' (transform failed "
+        f"or dispatch key drifted on this TRL)"
+    )
+    # The transformed body (__init__ rewrites, injected per-token-logps) lives in
+    # the generated module's `_UnslothGRPOTrainer` base + module-level funcs, not
+    # the thin UnslothGRPOTrainer subclass -- read the whole generated module.
+    mod = inspect.getmodule(patched)
+    return inspect.getsource(mod) if mod is not None else inspect.getsource(patched)
+
+
+@pytest.fixture(scope = "module")
+def generated_grpo_source():
+    if importlib.util.find_spec("unsloth") is None:
+        pytest.skip("unsloth not installed")
+    if importlib.util.find_spec("trl") is None:
+        pytest.skip("trl not installed")
+    # Do NOT swallow import errors: unsloth is installed here, so a failing
+    # `import unsloth` is exactly the import-time TRL/transformers drift this
+    # canary must surface as a failure, not a skip.
+    import unsloth  # noqa: F401  -- _gpu_init bootstrap under spoof
+
+    return _patch_grpo_and_get_source()
+
+
+def test_grpo_patch_generates_valid_source(generated_grpo_source):
+    """The generated UnslothGRPOTrainer must be syntactically valid Python."""
+    ast.parse(generated_grpo_source)
+
+
+def test_grpo_patch_aux_fail_fast_injected(generated_grpo_source):
+    """TRL >= 1.7.0: rl.py injects a fail-fast for the unsupported MoE router
+    aux-loss opt-in right after `self.aux_loss_enabled = ...`."""
+    from packaging.version import Version
+
+    if _trl_version() < Version("1.7.0"):
+        pytest.skip("aux_loss_enabled / router_aux_loss_coef are TRL >= 1.7.0")
+    assert "does not compute the MoE router auxiliary loss" in generated_grpo_source, (
+        "aux fail-fast raise missing from generated trainer; rl.py's "
+        "aux_loss_enabled .replace() anchor did not match this TRL"
+    )
+
+
+def test_grpo_patch_three_tuple_return(generated_grpo_source):
+    """TRL >= 1.7.0 call sites unpack a 3-tuple from
+    _get_per_token_logps_and_entropies; the injected replacement must return
+    (logps, entropies, aux_loss)."""
+    from packaging.version import Version
+    if _trl_version() >= Version("1.7.0"):
+        assert "return logprobs.detach(), entropies, aux_loss" in generated_grpo_source, (
+            "3-tuple per-token-logps return missing; the arity version-gate in "
+            "rl_replacements.py did not emit the >=1.7.0 form"
+        )
+    else:
+        assert (
+            "return logprobs.detach(), entropies, aux_loss" not in generated_grpo_source
+        ), "2-tuple TRL got the 3-tuple return; arity gate mis-fired"
+
+
+def test_grpo_patch_preserves_grad_checkpointing_block(generated_grpo_source):
+    """The tightened PR #6904 PEFT regex must remove only the ref-adapter init,
+    not the following enable_input_require_grads gradient-checkpointing block."""
+    from packaging.version import Version
+
+    if _trl_version() < Version("1.7.0"):
+        pytest.skip("ref-adapter elif block is the TRL >= 1.7.0 shape")
+    assert "enable_input_require_grads" in generated_grpo_source, (
+        "gradient-checkpointing enable_input_require_grads() block was swallowed "
+        "by the PEFT-removal regex (over-reach regression)"
+    )
+
+
+def test_grpo_patch_neutralizes_ref_adapter_and_qlora_cast(generated_grpo_source):
+    """TRL >= 1.7.0: the ref-adapter copy and the hardcoded QLoRA bf16 cast must
+    both be gone from the generated trainer."""
+    from packaging.version import Version
+
+    if _trl_version() < Version("1.7.0"):
+        pytest.skip("targets the TRL >= 1.7.0 PEFT / _is_quantized_model shapes")
+    assert (
+        "ref_param.data.copy_(param.data)" not in generated_grpo_source
+    ), "TRL's PEFT ref-adapter init survived; rl.py peft_pattern re.sub no-oped"
+    assert (
+        "if _is_quantized_model:" not in generated_grpo_source
+    ), "TRL's hardcoded QLoRA bf16 cast survived; rl.py neutralization no-oped"
+
+
+# SFT / DPO: the same source-transform patcher runs on them (a fake patch run,
+# no training), so a structural TRL change can break generation. Assert the patch
+# produces a valid, importable Unsloth trainer AND that the shared QLoRA
+# `_is_quantized_model` bf16 cast is neutralized (TRL 1.7's spelling), which the
+# patcher applies to every trainer. Catches "and or others" beyond GRPO.
+
+
+def _patch_and_get_source(trainer_file: str, trainer_cls: str) -> str:
+    if importlib.util.find_spec("unsloth") is None or importlib.util.find_spec("trl") is None:
+        pytest.skip("unsloth or trl not installed")
+    # Let a real import failure fail the test (import-time drift is the target).
+    import unsloth  # noqa: F401
+    import trl.trainer  # noqa: F401
+
+    from unsloth.models import rl as _rl
+
+    _rl._patch_trl_rl_trainers_impl(trainer_file)
+    mod = importlib.import_module(f"trl.trainer.{trainer_file}")
+    patched = getattr(mod, trainer_cls)
+    assert patched.__name__ == f"Unsloth{trainer_cls}", (
+        f"{trainer_cls} patch silently no-oped on this TRL "
+        f"(got {patched.__name__!r}); source-transform dispatch drifted"
+    )
+    gen = inspect.getmodule(patched)
+    src = inspect.getsource(gen) if gen is not None else inspect.getsource(patched)
+    ast.parse(src)
+    return src
+
+
+def _assert_quantized_cast_neutralized(src: str, trainer_cls: str) -> None:
+    from packaging.version import Version
+    if _trl_version() < Version("1.7.0"):
+        pytest.skip("pre-1.7.0 spells the QLoRA cast differently (is_loaded_in_4bit)")
+    assert "if _is_quantized_model:" not in src, (
+        f"{trainer_cls}: TRL's hardcoded QLoRA bf16 cast survived; the shared "
+        f"rl.py `if _is_quantized_model:` -> `if False:` neutralization no-oped"
+    )
+
+
+def test_sft_patch_generates_valid_source():
+    src = _patch_and_get_source("sft_trainer", "SFTTrainer")
+    _assert_quantized_cast_neutralized(src, "SFTTrainer")
+
+
+def test_dpo_patch_generates_valid_source():
+    src = _patch_and_get_source("dpo_trainer", "DPOTrainer")
+    _assert_quantized_cast_neutralized(src, "DPOTrainer")
+
+
+# The installed TRL in CI is always >= 1.7.0, so the < 1.7.0 return-arity
+# downgrade is never exercised by the fake-run above. Lock both arities by
+# monkeypatching rl_replacements.trl_version and re-generating the injected
+# _get_per_token_logps_and_entropies source directly (no TRL install needed).
+def test_per_token_logps_arity_gate_both_directions(monkeypatch):
+    if importlib.util.find_spec("unsloth") is None:
+        pytest.skip("unsloth not installed")
+    import unsloth  # noqa: F401
+    from packaging.version import Version
+
+    from unsloth.models import rl_replacements as _rlr
+
+    gate = _rlr.grpo_trainer__get_per_token_logps_and_entropies
+
+    # >= 1.7.0: 3-tuple return kept.
+    monkeypatch.setattr(_rlr, "trl_version", Version("1.7.0"), raising = False)
+    src_new = gate("_get_per_token_logps_and_entropies", None)
+    assert (
+        "return logprobs.detach(), entropies, aux_loss" in src_new
+    ), "3-tuple return missing for TRL >= 1.7.0"
+
+    # < 1.7.0: aux_loss element dropped -> 2-tuple. A no-op downgrade must raise
+    # (fail loud), never silently ship a 3-tuple to older TRL.
+    monkeypatch.setattr(_rlr, "trl_version", Version("1.6.0"), raising = False)
+    src_old = gate("_get_per_token_logps_and_entropies", None)
+    assert (
+        "return logprobs.detach(), entropies  # logps, entropies" in src_old
+    ), "2-tuple return missing for TRL < 1.7.0"
+    assert (
+        "entropies, aux_loss" not in src_old
+    ), "aux_loss element still present in the TRL < 1.7.0 downgrade"
diff --git a/tests/version_compat/test_trl_grpo_pinned_symbols.py b/tests/version_compat/test_trl_grpo_pinned_symbols.py
index 834692e2dd..2f935dde49 100644
--- a/tests/version_compat/test_trl_grpo_pinned_symbols.py
+++ b/tests/version_compat/test_trl_grpo_pinned_symbols.py
@@ -51,10 +51,27 @@ TRL_TAGS = [
     "v1.2.0",
     "v1.3.0",
     "v1.4.0",
+    "v1.5.0",
+    "v1.5.1",
+    "v1.6.0",
+    "v1.7.0",  # anchor: first release unsloth's TRL>=1.7.0 GRPO patch targets
+    "v1.7.1",  # current PyPI latest
     "main",
 ]
 
 
+def _tag_ge(tag: str, floor: str) -> bool:
+    """True if `tag` is `main` or a version >= `floor` (e.g. "1.7.0")."""
+    if tag == "main":
+        return True
+    from packaging.version import Version
+
+    try:
+        return Version(tag.lstrip("v")) >= Version(floor)
+    except Exception:
+        return False
+
+
 # unsloth/trainer.py + unsloth/models/rl.py rebind these top-level names.
 
 
@@ -537,3 +554,96 @@ def test_trl_truncate_with_protected_tokens_optional(tag: str):
     assert src is not None
     has_it = "truncate_with_protected_tokens" in src
     _ = has_it  # informational; pass either way.
+
+
+# 24-27. TRL >= 1.7.0 GRPO source contracts. Unlike the has_def existence
+# checks above, these pin the exact source strings unsloth/models/rl.py and
+# rl_replacements.py transform for TRL >= 1.7.0 (the window PR #6904 fixes).
+# The 1.7.0 break was invisible to the existence checks because the methods
+# still existed -- only their internal structure / return arity changed. If
+# TRL restructures one of these, the transform silently no-ops (or the
+# generated trainer breaks), so failing here on `main` gives a few-day lead.
+
+
+@pytest.mark.parametrize("tag", TRL_TAGS)
+def test_trl_grpo_peft_ref_adapter_block_contract(tag: str):
+    """rl.py (trl>=1.4.0) strips TRL's PEFT ref-adapter init with a re.DOTALL
+    regex anchored on `elif is_peft_model(model) and args.beta != 0.0:` ...
+    `ref_param.data.copy_(param.data)`. Both anchors must exist (else the
+    regex no-ops and the ref adapter is created under Unsloth), and the
+    following `enable_input_require_grads` gradient-checkpointing block must
+    remain present -- the tightened regex must NOT swallow it (PR #6904). The
+    `elif` block shape appeared in TRL 1.4.0, so this contract runs from there."""
+    if not _tag_ge(tag, "1.4.0"):
+        pytest.skip(
+            f"{tag}: pre-1.4.0 uses the `if is_peft_available()...` form (rl.py 0.27 branch)"
+        )
+    src = fetch_text("huggingface/trl", tag, "trl/trainer/grpo_trainer.py")
+    assert src is not None
+    assert "elif is_peft_model(model) and args.beta != 0.0:" in src, (
+        f"{tag}: PEFT ref-adapter `elif` anchor gone; unsloth/models/rl.py "
+        f"peft_pattern re.sub no-ops and TRL's ref adapter init runs under Unsloth"
+    )
+    assert "ref_param.data.copy_(param.data)" in src, (
+        f"{tag}: `ref_param.data.copy_(param.data)` end-anchor gone; "
+        f"unsloth/models/rl.py peft_pattern loses its DOTALL end match"
+    )
+    assert "enable_input_require_grads" in src, (
+        f"{tag}: `enable_input_require_grads` block gone from grpo_trainer.py; "
+        f"the tightened PR #6904 regex assumed it follows the ref-adapter block"
+    )
+
+
+@pytest.mark.parametrize("tag", TRL_TAGS)
+def test_trl_grpo_quantized_model_cast_contract(tag: str):
+    """rl.py (trl>=1.7.0) neutralizes TRL's hardcoded QLoRA bf16 cast
+    `if _is_quantized_model:` -> `if False:`. A rename leaves the cast active,
+    which ignores the user's dtype and breaks GradScaler with fp16=True."""
+    if not _tag_ge(tag, "1.7.0"):
+        pytest.skip(f"{tag}: pre-1.7.0 spells the cast differently (is_loaded_in_4bit)")
+    src = fetch_text("huggingface/trl", tag, "trl/trainer/grpo_trainer.py")
+    assert src is not None
+    assert "if _is_quantized_model:" in src, (
+        f"{tag}: `if _is_quantized_model:` gone; unsloth/models/rl.py cannot "
+        f"neutralize TRL's hardcoded QLoRA bf16 cast and it runs under Unsloth"
+    )
+
+
+@pytest.mark.parametrize("tag", TRL_TAGS)
+def test_trl_grpo_aux_loss_enabled_contract(tag: str):
+    """rl.py (trl>=1.7.0) appends a fail-fast after
+    `self.aux_loss_enabled = is_moe and args.router_aux_loss_coef != 0.0` so an
+    explicit MoE router-aux opt-in errors instead of silently training without
+    the penalty (the optimized forward cannot compute it). A change to this
+    line drops the guard silently (PR #6904)."""
+    if not _tag_ge(tag, "1.7.0"):
+        pytest.skip(f"{tag}: aux_loss_enabled / router_aux_loss_coef added in TRL 1.7.0")
+    src = fetch_text("huggingface/trl", tag, "trl/trainer/grpo_trainer.py")
+    assert src is not None
+    assert "self.aux_loss_enabled = is_moe and args.router_aux_loss_coef != 0.0" in src, (
+        f"{tag}: `aux_loss_enabled = is_moe and args.router_aux_loss_coef != 0.0` "
+        f"changed; unsloth/models/rl.py's fail-fast .replace() anchor no-ops"
+    )
+
+
+@pytest.mark.parametrize("tag", TRL_TAGS)
+def test_trl_grpo_per_token_logps_aux_arity_contract(tag: str):
+    """TRL 1.7.0 added `compute_aux_loss` to
+    _get_per_token_logps_and_entropies and made every call site unpack a
+    3-tuple. rl_replacements.py version-gates its injected replacement to emit
+    a 3-tuple for trl>=1.7.0 (2-tuple below). This is the exact change the
+    has_def existence checks miss: the method still exists, only its arity
+    changed. If TRL drops/renames the aux return, the gate needs revisiting."""
+    if not _tag_ge(tag, "0.20.0"):
+        pytest.skip(f"{tag}: pre-0.20 uses legacy _get_per_token_logps (2-tuple, no aux)")
+    src = fetch_text("huggingface/trl", tag, "trl/trainer/grpo_trainer.py")
+    assert src is not None
+    assert has_def(src, "_get_per_token_logps_and_entropies", "func"), (
+        f"{tag}: _get_per_token_logps_and_entropies missing on TRL >=0.20; "
+        f"unsloth's per-token-logps injection dispatch key no longer matches"
+    )
+    if _tag_ge(tag, "1.7.0"):
+        assert "compute_aux_loss" in src, (
+            f"{tag}: TRL >=1.7.0 dropped `compute_aux_loss`; the 3-tuple "
+            f"injection gate in unsloth/models/rl_replacements.py must be revisited"
+        )
diff --git a/unsloth-cli.py b/unsloth-cli.py
index 756efef0d0..be48893d8d 100644
--- a/unsloth-cli.py
+++ b/unsloth-cli.py
@@ -22,24 +22,179 @@ import argparse
 import os
 
 
+def _is_mlx_backend(unsloth_module):
+    return bool(getattr(unsloth_module, "_IS_MLX", False))
+
+
+def _normalize_dtype(dtype, is_mlx):
+    if is_mlx and isinstance(dtype, str) and dtype.strip().lower() in {"", "none", "auto"}:
+        return None
+    return dtype
+
+
+def _prepare_device_map(is_mlx):
+    if is_mlx:
+        return None, False
+
+    from unsloth.models.loader_utils import prepare_device_map
+    return prepare_device_map()
+
+
+class _CallableTokenizerProxy:
+    def __init__(self, tokenizer):
+        self._tokenizer = tokenizer
+
+    def __getattr__(self, name):
+        return getattr(self._tokenizer, name)
+
+    def __call__(self, text, *args, **kwargs):
+        # MLX/torch-free: never request torch tensors; keep plain python ids.
+        kwargs.pop("return_tensors", None)
+        wrapped = getattr(self._tokenizer, "_tokenizer", None)
+        if callable(wrapped):
+            return wrapped(text, *args, **kwargs)
+
+        add_special_tokens = kwargs.get("add_special_tokens", False)
+        input_ids = self._tokenizer.encode(text, add_special_tokens = add_special_tokens)
+        return {"input_ids": input_ids}
+
+
+def _tokenizer_for_raw_text_loader(tokenizer, is_mlx):
+    if not is_mlx or callable(tokenizer):
+        return tokenizer
+    return _CallableTokenizerProxy(tokenizer)
+
+
+def _raw_text_loader_for_backend(
+    RawTextDataLoader,
+    tokenizer,
+    is_mlx,
+    chunk_size = 2048,
+    stride = 512,
+):
+    return RawTextDataLoader(
+        _tokenizer_for_raw_text_loader(tokenizer, is_mlx),
+        chunk_size,
+        stride,
+        return_tokenized = not is_mlx,
+    )
+
+
+def _train_with_legacy_save_control(trainer, is_mlx):
+    if not is_mlx:
+        return trainer.train()
+
+    original_save_model = getattr(trainer, "save_model", None)
+    if original_save_model is None:
+        return trainer.train()
+
+    def skip_internal_final_save(*args, **kwargs):
+        raise ValueError("legacy unsloth-cli.py owns final save")
+
+    trainer.save_model = skip_internal_final_save
+    try:
+        return trainer.train()
+    finally:
+        trainer.save_model = original_save_model
+
+
+def _iter_quantization_methods(quantization):
+    if isinstance(quantization, list):
+        return quantization
+    return [quantization]
+
+
+def _save_or_push_model(model, tokenizer, args, is_mlx):
+    if not args.save_model:
+        print("Warning: The model is not saved!")
+        return
+
+    # Enter the GGUF branch when saving or pushing GGUF, so --push_gguf works
+    # without --save_gguf (the local save is guarded separately below).
+    if args.save_gguf or args.push_gguf:
+        if not args.save_gguf:
+            print("Warning: --save_gguf not set, pushing GGUF to hub without saving locally.")
+        for quantization_method in _iter_quantization_methods(args.quantization):
+            if args.save_gguf:
+                print(f"Saving model with quantization method: {quantization_method}")
+                model.save_pretrained_gguf(
+                    args.save_path,
+                    tokenizer,
+                    quantization_method = quantization_method,
+                )
+            if args.push_model or args.push_gguf:
+                model.push_to_hub_gguf(
+                    args.hub_path,
+                    tokenizer,
+                    quantization_method = quantization_method,
+                    token = args.hub_token,
+                )
+        return
+
+    if is_mlx:
+        model.save_pretrained_merged(
+            args.save_path,
+            tokenizer,
+            save_method = args.save_method,
+            push_to_hub = args.push_model,
+            repo_id = args.hub_path if args.push_model else None,
+            token = args.hub_token,
+        )
+        return
+
+    model.save_pretrained_merged(args.save_path, tokenizer, save_method = args.save_method)
+    if args.push_model:
+        model.push_to_hub_merged(args.hub_path, tokenizer, args.save_method, token = args.hub_token)
+
+
+def _build_sft_config(SFTConfig, args, is_mlx, bf16_supported):
+    config_kwargs = dict(
+        per_device_train_batch_size = args.per_device_train_batch_size,
+        gradient_accumulation_steps = args.gradient_accumulation_steps,
+        warmup_steps = args.warmup_steps,
+        max_steps = args.max_steps,
+        learning_rate = args.learning_rate,
+        fp16 = not bf16_supported,
+        bf16 = bf16_supported,
+        logging_steps = args.logging_steps,
+        optim = args.optim,
+        weight_decay = args.weight_decay,
+        lr_scheduler_type = args.lr_scheduler_type,
+        seed = args.seed,
+        output_dir = args.output_dir,
+        report_to = args.report_to,
+        max_length = args.max_seq_length,
+        dataset_num_proc = 2,
+        packing = args.packing,
+    )
+    if is_mlx:
+        if args.per_device_eval_batch_size != 4:
+            print("Warning: --per_device_eval_batch_size is ignored on MLX without eval data.")
+    else:
+        config_kwargs["per_device_eval_batch_size"] = args.per_device_eval_batch_size
+    return SFTConfig(**config_kwargs)
+
+
 def run(args):
+    import unsloth
     from unsloth import FastLanguageModel
     from datasets import load_dataset
     from transformers.utils import strtobool
     from trl import SFTTrainer, SFTConfig
     from unsloth import is_bfloat16_supported
-    from unsloth.models.loader_utils import prepare_device_map
     import logging
     from unsloth import RawTextDataLoader
 
     logging.getLogger("hf-to-gguf").setLevel(logging.WARNING)
 
+    is_mlx = _is_mlx_backend(unsloth)
+
     # Load model and tokenizer
-    device_map, distributed = prepare_device_map()
+    device_map, distributed = _prepare_device_map(is_mlx)
     model, tokenizer = FastLanguageModel.from_pretrained(
         model_name = args.model_name,
         max_seq_length = args.max_seq_length,
-        dtype = args.dtype,
+        dtype = _normalize_dtype(args.dtype, is_mlx),
         load_in_4bit = args.load_in_4bit,
         device_map = device_map,
     )
@@ -92,11 +247,13 @@ def run(args):
     def load_dataset_smart(args):
         from transformers.utils import strtobool
         if args.raw_text_file:
-            loader = RawTextDataLoader(tokenizer, args.chunk_size, args.stride)
+            loader = _raw_text_loader_for_backend(
+                RawTextDataLoader, tokenizer, is_mlx, args.chunk_size, args.stride
+            )
             dataset = loader.load_from_file(args.raw_text_file)
         elif args.dataset.endswith((".txt", ".md", ".json", ".jsonl")):
             # Auto-detect local raw text files
-            loader = RawTextDataLoader(tokenizer)
+            loader = _raw_text_loader_for_backend(RawTextDataLoader, tokenizer, is_mlx)
             dataset = loader.load_from_file(args.dataset)
         else:
             use_modelscope = strtobool(os.environ.get("UNSLOTH_USE_MODELSCOPE", "False"))
@@ -115,27 +272,9 @@ def run(args):
     print("Data is formatted and ready!")
 
     # Configure training arguments
-    training_args = SFTConfig(
-        per_device_train_batch_size = args.per_device_train_batch_size,
-        per_device_eval_batch_size = args.per_device_eval_batch_size,
-        gradient_accumulation_steps = args.gradient_accumulation_steps,
-        warmup_steps = args.warmup_steps,
-        max_steps = args.max_steps,
-        learning_rate = args.learning_rate,
-        fp16 = not is_bfloat16_supported(),
-        bf16 = is_bfloat16_supported(),
-        logging_steps = args.logging_steps,
-        optim = args.optim,
-        weight_decay = args.weight_decay,
-        lr_scheduler_type = args.lr_scheduler_type,
-        seed = args.seed,
-        output_dir = args.output_dir,
-        report_to = args.report_to,
-        max_length = args.max_seq_length,
-        dataset_num_proc = 2,
-        ddp_find_unused_parameters = False if distributed else None,
-        packing = args.packing,
-    )
+    training_args = _build_sft_config(SFTConfig, args, is_mlx, is_bfloat16_supported())
+    if distributed:
+        training_args.ddp_find_unused_parameters = False
 
     # Initialize trainer
     trainer = SFTTrainer(
@@ -145,57 +284,9 @@ def run(args):
         args = training_args,
     )
 
-    trainer.train()
+    _train_with_legacy_save_control(trainer, is_mlx)
 
-    # Save model
-    if args.save_model:
-        # If args.quantization is a list, save once per quantization method
-        # Enter the GGUF branch when saving *or* pushing GGUF, so --push_gguf
-        # works even when --save_gguf is omitted (the local save is guarded
-        # separately below).
-        if args.save_gguf or args.push_gguf:
-            # Push-only GGUF (no --save_gguf) skips the local save; warn so it is not silent.
-            if not args.save_gguf:
-                print("Warning: --save_gguf not set, pushing GGUF to hub without saving locally.")
-            if isinstance(args.quantization, list):
-                for quantization_method in args.quantization:
-                    if args.save_gguf:
-                        print(f"Saving model with quantization method: {quantization_method}")
-                        model.save_pretrained_gguf(
-                            args.save_path,
-                            tokenizer,
-                            quantization_method = quantization_method,
-                        )
-                    if args.push_model or args.push_gguf:
-                        model.push_to_hub_gguf(
-                            args.hub_path,
-                            tokenizer,
-                            quantization_method = quantization_method,
-                            token = args.hub_token,
-                        )
-            else:
-                if args.save_gguf:
-                    print(f"Saving model with quantization method: {args.quantization}")
-                    model.save_pretrained_gguf(
-                        args.save_path,
-                        tokenizer,
-                        quantization_method = args.quantization,
-                    )
-                if args.push_model or args.push_gguf:
-                    model.push_to_hub_gguf(
-                        args.hub_path,
-                        tokenizer,
-                        quantization_method = args.quantization,
-                        token = args.hub_token,
-                    )
-        else:
-            model.save_pretrained_merged(args.save_path, tokenizer, args.save_method)
-            if args.push_model:
-                model.push_to_hub_merged(
-                    args.hub_path, tokenizer, args.save_method, token = args.hub_token
-                )
-    else:
-        print("Warning: The model is not saved!")
+    _save_or_push_model(model, tokenizer, args, is_mlx)
 
 
 if __name__ == "__main__":
diff --git a/unsloth/__init__.py b/unsloth/__init__.py
index 04cc600725..de5cd0f61b 100644
--- a/unsloth/__init__.py
+++ b/unsloth/__init__.py
@@ -102,6 +102,7 @@ if _IS_MLX:
         ) from _e
 
     import dataclasses as _dataclasses
+    import inspect as _inspect
     import importlib.machinery as _machinery
     import sys as _sys
     import types as _types
@@ -109,6 +110,30 @@ if _IS_MLX:
 
     __version__ = unsloth_zoo.__version__
     DEVICE_TYPE = "mlx"
+    _MLX_TRAINER_ACCEPTS_VAR_KWARGS = False
+    _MLX_TRAINER_SUPPORTED_KWARGS = frozenset()
+    try:
+        _MLX_TRAINER_INIT_PARAMETERS = _inspect.signature(MLXTrainer.__init__).parameters
+        _MLX_TRAINER_ACCEPTS_VAR_KWARGS = any(
+            param.kind is _inspect.Parameter.VAR_KEYWORD
+            for param in _MLX_TRAINER_INIT_PARAMETERS.values()
+        )
+        _MLX_TRAINER_SUPPORTED_KWARGS = frozenset(
+            name
+            for name, param in _MLX_TRAINER_INIT_PARAMETERS.items()
+            if name != "self"
+            and param.kind
+            in (
+                _inspect.Parameter.POSITIONAL_OR_KEYWORD,
+                _inspect.Parameter.KEYWORD_ONLY,
+            )
+        )
+    except (TypeError, ValueError):
+        pass
+
+    def _mlx_trainer_supports_kwarg(name):
+        """Return whether the installed zoo MLXTrainer accepts a kwarg."""
+        return _MLX_TRAINER_ACCEPTS_VAR_KWARGS or name in _MLX_TRAINER_SUPPORTED_KWARGS
 
     def _is_mlx_cuda_device_target(device):
         """Return True when a torch .to/.cuda target asks for CUDA on MLX."""
@@ -966,6 +991,7 @@ if _IS_MLX:
         "args",
         "formatting_func",
         "processor",
+        "callbacks",
     )
     _TRL_SFT_TRAINER_POSITIONAL_KWARGS = (
         "model",
@@ -985,6 +1011,29 @@ if _IS_MLX:
     )
     _MLX_TRAINER_KWARGS = frozenset(_MLX_TRAINER_POSITIONAL_KWARGS)
 
+    def _filter_supported_mlx_trainer_kwargs(trainer_kwargs):
+        """Drop inert/empty kwargs unsupported by this zoo MLXTrainer."""
+        unsupported = {
+            key: value
+            for key, value in trainer_kwargs.items()
+            if not _mlx_trainer_supports_kwarg(key)
+        }
+        names = sorted(
+            key for key, value in unsupported.items() if _is_meaningful_mlx_extra_value(value)
+        )
+        if names:
+            subject = ", ".join(names)
+            verb = "requires" if len(names) == 1 else "require"
+            raise NotImplementedError(
+                "Unsloth MLX: "
+                f"{subject} {verb} an unsloth-zoo build with "
+                "matching MLXTrainer support. Upgrade unsloth-zoo together "
+                "with unsloth."
+            )
+        for key in unsupported:
+            trainer_kwargs.pop(key, None)
+        return trainer_kwargs
+
     def _is_mlx_native_text_collator(collator):
         """HF pad/copy collators are redundant on MLX; match by class name."""
         for klass in type(collator).__mro__:
@@ -1162,6 +1211,7 @@ if _IS_MLX:
 
             trainer_kwargs, config_kwargs, ignored_kwargs = _split_mlx_trainer_kwargs(kwargs)
             _raise_unsupported_mlx_trainer_kwargs(ignored_kwargs)
+            trainer_kwargs = _filter_supported_mlx_trainer_kwargs(trainer_kwargs)
             trainer_kwargs["args"] = _coerce_mlx_training_args(
                 trainer_kwargs.get("args"),
                 config_kwargs,
diff --git a/unsloth/_gpu_init.py b/unsloth/_gpu_init.py
index e39b44488a..984057e9f7 100644
--- a/unsloth/_gpu_init.py
+++ b/unsloth/_gpu_init.py
@@ -26,6 +26,7 @@ already_imported = [mod for mod in critical_modules if mod in sys.modules]
 # Fix some issues before importing other packages
 from .import_fixes import (
     fix_message_factory_issue,
+    fix_torch_check_is_size,
     check_fbgemm_gpu_version,
     disable_broken_causal_conv1d,
     disable_broken_vllm,
@@ -72,6 +73,7 @@ fix_bitsandbytes_rocm_arch_detection()
 disable_broken_causal_conv1d()
 disable_broken_vllm()
 fix_message_factory_issue()
+fix_torch_check_is_size()
 check_fbgemm_gpu_version()
 torchvision_compatibility_check()
 fix_diffusers_warnings()
@@ -81,6 +83,7 @@ del fix_bitsandbytes_rocm_arch_detection
 del disable_broken_causal_conv1d
 del disable_broken_vllm
 del fix_message_factory_issue
+del fix_torch_check_is_size
 del check_fbgemm_gpu_version
 del torchvision_compatibility_check
 del fix_diffusers_warnings
@@ -173,6 +176,7 @@ from .import_fixes import (
     fix_vllm_guided_decoding_params,
     fix_vllm_pdl_blackwell,
     fix_triton_compiled_kernel_missing_attrs,
+    fix_dynamo_config_thread_visibility,
     patch_trunc_normal_precision_issue,
     ignore_logger_messages,
     patch_ipykernel_hf_xet,
@@ -203,6 +207,10 @@ fix_vllm_guided_decoding_params()
 fix_trl_vllm_ascend()
 fix_vllm_pdl_blackwell()
 fix_triton_compiled_kernel_missing_attrs()
+# Must run before unsloth_zoo's patch_torch_compile and the gpt-oss temporary
+# patches raise the dynamo recompile limits, so those settings reach the
+# autograd worker threads on torch >= 2.12.
+fix_dynamo_config_thread_visibility()
 patch_trunc_normal_precision_issue()
 ignore_logger_messages()
 patch_ipykernel_hf_xet()
@@ -233,6 +241,7 @@ del fix_vllm_guided_decoding_params
 del fix_trl_vllm_ascend
 del fix_vllm_pdl_blackwell
 del fix_triton_compiled_kernel_missing_attrs
+del fix_dynamo_config_thread_visibility
 del patch_trunc_normal_precision_issue
 del ignore_logger_messages
 del patch_ipykernel_hf_xet
diff --git a/unsloth/chat_templates.py b/unsloth/chat_templates.py
index 169b2dbd0e..2d3674fb04 100644
--- a/unsloth/chat_templates.py
+++ b/unsloth/chat_templates.py
@@ -2103,8 +2103,9 @@ def get_chat_template(
 
 def remove_special_tokens(tokenizer, prompt):
     # Removes double BOS token
-    if prompt.startswith(tokenizer.bos_token):
-        prompt = prompt[len(tokenizer.bos_token):]
+    bos_token = getattr(tokenizer, "bos_token", None)
+    if bos_token is not None and prompt.startswith(bos_token):
+        prompt = prompt[len(bos_token):]
     return prompt
 
 
@@ -2185,7 +2186,16 @@ def _create_formatter(possible_columns, final_optional_prompts, user_column_name
 
         texts = []
         for row_idx in range(n_rows):
-            row_values = {column: examples[column][row_idx] for column in columns}
+            # Coerce missing (None) columns to "" so they do not render as the
+            # literal string "None" in the emitted text. In a [[...]] block only
+            # the first column gates the block, so a later column can still be
+            # None here; required columns can be None too. Coercing at the source
+            # covers both; since None is now "", the gate below only needs to
+            # test for "" (an empty first column still drops the block).
+            row_values = {
+                column: ("" if (value := examples[column][row_idx]) is None else value)
+                for column in columns
+            }
             formatter_values = {}
 
             for formatter_template in formatter_templates:
@@ -2196,7 +2206,7 @@ def _create_formatter(possible_columns, final_optional_prompts, user_column_name
                     continue
 
                 _, optional_name, prompt, needed_columns = formatter_template
-                if row_values[needed_columns[0]] not in (None, ""):
+                if row_values[needed_columns[0]] != "":
                     prompt_values = {column: row_values[column] for column in needed_columns}
                     formatter_values[optional_name] = prompt.format(**prompt_values)
                 else:
diff --git a/unsloth/import_fixes.py b/unsloth/import_fixes.py
index e5a5d01c2f..09de248c7b 100644
--- a/unsloth/import_fixes.py
+++ b/unsloth/import_fixes.py
@@ -172,6 +172,10 @@ if not UNSLOTH_ENABLE_LOGGING:
     # Deprecation warnings from torchao
     warnings.filterwarnings("ignore", message = "`int4_weight_only` is deprecated")
     warnings.filterwarnings("ignore", message = "`int8_weight_only` is deprecated")
+    # torch._check_is_size FutureWarning (called by bitsandbytes 4-bit dequant)
+    warnings.filterwarnings(
+        "ignore", message = r"_check_is_size will be removed", category = FutureWarning
+    )
 
     # TorchAO deprecated import paths (https://github.com/pytorch/ao/issues/2752)
     warnings.filterwarnings(
@@ -253,6 +257,30 @@ if not UNSLOTH_ENABLE_LOGGING:
     )
 
 
+def fix_torch_check_is_size():
+    """Shim torch._check_is_size if a future torch removes it (bitsandbytes 4-bit
+    dequant calls it). The FutureWarning is silenced in suppress_cuda_printf."""
+    try:
+        import torch
+
+        if hasattr(torch, "_check_is_size"):
+            return
+
+        def _check_is_size(
+            i,
+            message = None,
+            *,
+            max = None,
+        ):
+            torch._check(i >= 0, message)
+            if max is not None:
+                torch._check(i <= max, message)
+
+        torch._check_is_size = _check_is_size
+    except Exception:
+        return
+
+
 # Fix up AttributeError: 'MessageFactory' object has no attribute 'GetPrototype'
 # MUST do this at the start primarily due to tensorflow causing issues
 def fix_message_factory_issue():
@@ -1064,6 +1092,135 @@ def fix_triton_compiled_kernel_missing_attrs():
     )
 
 
+def fix_dynamo_config_thread_visibility():
+    """torch 2.12 made torch._dynamo/_inductor config overrides thread-local
+    (ContextVars), so `config.recompile_limit = 1024` set on the main thread is
+    invisible to the autograd worker threads that run backward. Gradient
+    checkpointing recompiles fullgraph gpt-oss kernels there against the default
+    limit of 8, raising FailOnRecompileLimitHit at step 0. Mirror direct config
+    assignments into the process-global entry default (torch <= 2.11 semantics).
+    config.patch(...) and config.load_config(...) also assign via __setattr__ but
+    are thread-local by design, so skip mirroring while inside one (tracked per
+    thread). No-op below torch 2.12 and on any torch without this internal layout.
+    """
+    try:
+        import torch
+
+        if Version(torch.__version__) < Version("2.12.0"):
+            return
+        import torch._dynamo.config as _dynamo_config
+        from torch.utils._config_module import ConfigModule
+        from contextvars import ContextVar
+    except Exception:
+        return
+
+    try:
+        probe = getattr(_dynamo_config, "_config", {}).get("recompile_limit", None)
+        if probe is None or not isinstance(getattr(probe, "user_override", None), ContextVar):
+            # Overrides are not context-local on this torch; nothing to fix.
+            return
+        original_setattr = ConfigModule.__setattr__
+        if getattr(original_setattr, "__unsloth_patched__", False):
+            return
+    except Exception:
+        return
+
+    mirrored_modules = ("torch._dynamo.config", "torch._inductor.config")
+
+    # config.patch(...) and config.load_config(...) also assign via __setattr__, but
+    # their writes are thread-local by design; a per-thread depth counter marks them
+    # so they are not mirrored into the process-global default.
+    import threading
+
+    _scoped_depth = threading.local()
+
+    def _in_scoped_write():
+        return getattr(_scoped_depth, "n", 0) > 0
+
+    def _bump(delta):
+        _scoped_depth.n = getattr(_scoped_depth, "n", 0) + delta
+
+    original_patch = ConfigModule.patch
+    if not getattr(original_patch, "__unsloth_patched__", False):
+
+        @functools.wraps(original_patch)
+        def _patched_patch(self, *args, **kwargs):
+            ctx = original_patch(self, *args, **kwargs)
+            try:
+                cls = type(ctx)  # patch() builds a fresh ConfigPatch class each call
+                if not getattr(cls, "__unsloth_patch_wrapped__", False):
+                    _enter0, _exit0 = cls.__enter__, cls.__exit__
+
+                    def _enter(s, _e = _enter0):
+                        _bump(1)
+                        try:
+                            return _e(s)
+                        finally:
+                            _bump(-1)
+
+                    def _exit(
+                        s,
+                        *a,
+                        _x = _exit0,
+                    ):
+                        _bump(1)
+                        try:
+                            return _x(s, *a)
+                        finally:
+                            _bump(-1)
+
+                    cls.__enter__, cls.__exit__ = _enter, _exit
+                    cls.__unsloth_patch_wrapped__ = True
+            except Exception:
+                pass
+            return ctx
+
+        _patched_patch.__unsloth_patched__ = True
+        ConfigModule.patch = _patched_patch
+
+    # load_config restores a saved config by calling setattr per key (thread-local).
+    original_load_config = getattr(ConfigModule, "load_config", None)
+    if callable(original_load_config) and not getattr(
+        original_load_config, "__unsloth_patched__", False
+    ):
+
+        @functools.wraps(original_load_config)
+        def _patched_load_config(self, *args, **kwargs):
+            _bump(1)
+            try:
+                return original_load_config(self, *args, **kwargs)
+            finally:
+                _bump(-1)
+
+        _patched_load_config.__unsloth_patched__ = True
+        ConfigModule.load_config = _patched_load_config
+
+    @functools.wraps(original_setattr)
+    def _patched_setattr(self, name, value):
+        original_setattr(self, name, value)
+        if _in_scoped_write():
+            return  # transient patch / load_config write: keep it thread-local
+        # Aliases (cache_size_limit -> recompile_limit) re-enter with the real name.
+        if self.__dict__.get("__name__", None) in mirrored_modules:
+            try:
+                entry = self.__dict__["_config"].get(name, None)
+                if entry is not None and entry.alias is None:
+                    entry.default = value
+            except Exception:
+                pass
+
+    _patched_setattr.__unsloth_patched__ = True
+    ConfigModule.__setattr__ = _patched_setattr
+
+    # No replay of existing overrides: unsloth installs this before it sets any
+    # dynamo/inductor config, so the wrapper mirrors every later assignment. Replaying
+    # would also bake a still-active config.patch override into the global default.
+    logger.info(
+        "Unsloth: Patched torch config modules so dynamo/inductor settings "
+        "(e.g. recompile_limit) apply across threads on torch >= 2.12."
+    )
+
+
 def patch_trunc_normal_precision_issue():
     """
     Patch torch.nn.init.trunc_normal_ for low precision tensors to run init in fp32.
@@ -1323,8 +1480,7 @@ def fix_vllm_pdl_blackwell():
 
     if patched:
         logger.info(
-            f"Unsloth: Applied PDL fix for SM100 ({sm100_gpu_name}) - "
-            f"patched: {', '.join(patched)}"
+            f"Unsloth: Applied PDL fix for SM100 ({sm100_gpu_name}) - patched: {', '.join(patched)}"
         )
     else:
         # Just set the env var - vLLM might be an older version without supports_pdl
diff --git a/unsloth/kernels/fp8.py b/unsloth/kernels/fp8.py
index 935ffbb447..7a57f91ce1 100644
--- a/unsloth/kernels/fp8.py
+++ b/unsloth/kernels/fp8.py
@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import os
+from contextlib import nullcontext
 import torch
 import torch.nn as nn
 import triton
@@ -24,6 +25,15 @@ from unsloth_zoo.temporary_patches.common import torch_compile
 
 torch_matmul = torch.matmul
 
+
+def _fp8_triton_device_context(tensor: torch.Tensor):
+    if tensor.device.type == "cuda" and torch.cuda.device_count() > 1:
+        return torch.cuda.device(tensor.device)
+    if tensor.device.type == "xpu" and hasattr(torch, "xpu") and torch.xpu.device_count() > 1:
+        return torch.xpu.device(tensor.device)
+    return nullcontext()
+
+
 try:
     from transformers.integrations.finegrained_fp8 import FP8Linear
 except:
@@ -32,6 +42,11 @@ except:
         "Unsloth: FP8 models need importing FP8Linear from `transformers.integrations.finegrained_fp8` but we don't see it."
     )
 
+try:
+    from transformers.integrations.finegrained_fp8 import FP8GroupedLinear
+except:
+    FP8GroupedLinear = None
+
 try:
     from transformers.integrations.fbgemm_fp8 import FbgemmFp8Linear
 except:
@@ -95,7 +110,8 @@ def weight_dequant_block(
         triton.cdiv(M, meta["BLOCK_SIZE"]),
         triton.cdiv(N, meta["BLOCK_SIZE"]),
     )
-    weight_dequant_kernel[grid](x, s, y, M, N, BLOCK_SIZE = block_size)
+    with _fp8_triton_device_context(x):
+        weight_dequant_kernel[grid](x, s, y, M, N, BLOCK_SIZE = block_size)
     return y
 
 
@@ -149,7 +165,8 @@ def act_quant(x: torch.Tensor, block_size: int = 128) -> tuple[torch.Tensor, tor
     def grid(meta):
         return (triton.cdiv(x.numel(), meta["BLOCK_SIZE"]),)
 
-    act_quant_kernel[grid](x, y, s, BLOCK_SIZE = block_size)
+    with _fp8_triton_device_context(x):
+        act_quant_kernel[grid](x, y, s, BLOCK_SIZE = block_size)
     return y, s
 
 
@@ -274,32 +291,33 @@ def w8a8_block_fp8_matmul_triton(
     def grid(META):
         return (triton.cdiv(M, META["BLOCK_SIZE_M"]) * triton.cdiv(N, META["BLOCK_SIZE_N"]),)
 
-    _w8a8_block_fp8_matmul[grid](
-        A,
-        B,
-        C,
-        As,
-        Bs,
-        M,
-        N,
-        K,
-        block_n,
-        block_k,
-        A.stride(-2),
-        A.stride(-1),
-        B.stride(1),
-        B.stride(0),
-        C.stride(-2),
-        C.stride(-1),
-        As.stride(-2),
-        As.stride(-1),
-        Bs.stride(1),
-        Bs.stride(0),
-        BLOCK_SIZE_M = BLOCK_SIZE_M,
-        BLOCK_SIZE_N = BLOCK_SIZE_N,
-        BLOCK_SIZE_K = BLOCK_SIZE_K,
-        GROUP_SIZE_M = 8,
-    )
+    with _fp8_triton_device_context(A):
+        _w8a8_block_fp8_matmul[grid](
+            A,
+            B,
+            C,
+            As,
+            Bs,
+            M,
+            N,
+            K,
+            block_n,
+            block_k,
+            A.stride(-2),
+            A.stride(-1),
+            B.stride(1),
+            B.stride(0),
+            C.stride(-2),
+            C.stride(-1),
+            As.stride(-2),
+            As.stride(-1),
+            Bs.stride(1),
+            Bs.stride(0),
+            BLOCK_SIZE_M = BLOCK_SIZE_M,
+            BLOCK_SIZE_N = BLOCK_SIZE_N,
+            BLOCK_SIZE_K = BLOCK_SIZE_K,
+            GROUP_SIZE_M = 8,
+        )
     return C
 
 
@@ -311,13 +329,14 @@ def torchao_block_matmul(
     block_size: tuple[int, int],
     output_dtype: torch.dtype = torch.bfloat16,
 ):
-    out = torchao_blockwise_gemm(
-        act_q.contiguous(),
-        act_scale.contiguous(),
-        weight_q.contiguous(),
-        weight_scale.contiguous(),
-        block_size = block_size[1],
-    )
+    with _fp8_triton_device_context(act_q):
+        out = torchao_blockwise_gemm(
+            act_q.contiguous(),
+            act_scale.contiguous(),
+            weight_q.contiguous(),
+            weight_scale.contiguous(),
+            block_size = block_size[1],
+        )
     return out.to(output_dtype)
 
 
@@ -540,7 +559,8 @@ class FP8_fbgemm_block_linear(torch.autograd.Function):
                     f"Weight shape {weight.shape} and scales shape {weight_scale.shape} is not compatible with block size {bs_n, bs_k}"
                 )
 
-        xq, xs = triton_quantize_fp8_block(X, bs_m, bs_n, None)
+        with _fp8_triton_device_context(X):
+            xq, xs = triton_quantize_fp8_block(X, bs_m, bs_n, None)
         # TODO: WARNING - diverges from baseline for high X values, producing
         # gibberish / high starting loss. Do not use until resolved; kept for a
         # future headstart.
@@ -673,3 +693,66 @@ if FbgemmFp8Linear is not None:
     FbgemmFp8Linear.forward = module_forward_patch(fbgemm_fp8_linear, "weight_scale")
 if FP8Linear is not None:
     FP8Linear.forward = module_forward_patch(fp8_block_quant_linear, "weight_scale_inv")
+
+# FP8GroupedLinear's fused grouped matmul has no autograd formula, so training
+# backward fails. In training, use a custom autograd Function: dequant the frozen
+# fp8 weight for a differentiable bmm, saving only the fp8 weight + scale and
+# unwrapping TP shards; eval keeps the fused kernel. Gate on self.training (not
+# is_grad_enabled) so the grad-checkpoint no-grad forward and its recompute match.
+if FP8GroupedLinear is not None:
+    _fp8_grouped_forward_orig = FP8GroupedLinear.forward
+
+    def _fp8_to_local(t):
+        dt = getattr(getattr(torch, "distributed", None), "tensor", None)
+        DTensor = getattr(dt, "DTensor", None) if dt is not None else None
+        return t.to_local() if DTensor is not None and isinstance(t, DTensor) else t
+
+    def _fp8_grouped_dequant(weight, scale_inv, block_size, dtype):
+        # Honor the layer's block size; weight_dequant would assume 128 and mis-scale.
+        if block_size is not None and len(block_size) == 2:
+            return _blockwise_weight_dequant_any_shape(weight, scale_inv.float(), block_size, dtype)
+        return weight_dequant(weight, scale_inv.float()).to(dtype)
+
+    class _FP8GroupedMM(torch.autograd.Function):
+        @staticmethod
+        def forward(ctx, x, weight, scale_inv, n_groups, block_size, bias):
+            weight, scale_inv = _fp8_to_local(weight), _fp8_to_local(scale_inv)
+            hidden = x.shape[-1]
+            W = _fp8_grouped_dequant(weight, scale_inv, block_size, x.dtype)
+            out_per = W.shape[0] // n_groups
+            xg = x.reshape(-1, n_groups, hidden).transpose(0, 1)
+            y = torch.bmm(xg, W.view(n_groups, out_per, hidden).transpose(1, 2))
+            y = y.transpose(0, 1).reshape(*x.shape[:-2], n_groups, out_per)
+            if bias is not None:
+                y = y + bias.view(n_groups, out_per)
+            ctx.save_for_backward(weight, scale_inv)
+            ctx.n_groups, ctx.out_per, ctx.x_shape = n_groups, out_per, x.shape
+            ctx.dtype, ctx.has_bias, ctx.block_size = x.dtype, bias is not None, block_size
+            return y
+
+        @staticmethod
+        def backward(ctx, grad_y):
+            weight, scale_inv = ctx.saved_tensors
+            ng, out_per, hidden = ctx.n_groups, ctx.out_per, ctx.x_shape[-1]
+            W = _fp8_grouped_dequant(weight, scale_inv, ctx.block_size, ctx.dtype).view(
+                ng, out_per, hidden
+            )
+            gy = grad_y.reshape(-1, ng, out_per).transpose(0, 1)
+            grad_x = torch.bmm(gy, W).transpose(0, 1).reshape(ctx.x_shape)
+            grad_bias = gy.sum(1).reshape(-1) if ctx.has_bias else None
+            return grad_x, None, None, None, None, grad_bias
+
+    def _fp8_grouped_forward(self, x):
+        if self.weight.element_size() > 1 or not self.training:
+            return _fp8_grouped_forward_orig(self, x)
+        bias = self.bias if self.has_bias else None
+        return _FP8GroupedMM.apply(
+            x,
+            self.weight,
+            self.weight_scale_inv,
+            self.n_groups,
+            getattr(self, "block_size", None),
+            bias,
+        )
+
+    FP8GroupedLinear.forward = _fp8_grouped_forward
diff --git a/unsloth/kernels/utils.py b/unsloth/kernels/utils.py
index 43ed198a4a..1b0b5ce12e 100644
--- a/unsloth/kernels/utils.py
+++ b/unsloth/kernels/utils.py
@@ -282,6 +282,21 @@ def QUANT_STATE(W):
     return getattr(W, "quant_state", None)
 
 
+# fp8 weight dtypes. A `weight_scale` / `weight_scale_inv` should only be treated as a
+# quant state when the weight itself is still fp8. compressed-tensors layers expose an
+# already-dequantized bf16 weight at forward time while keeping a `weight_scale` around;
+# reading that as a quant state routes a bf16 weight into the bitsandbytes fast_gemv /
+# fast_dequantize path, which then reads a missing `absmax` and crashes.
+_FP8_WEIGHT_DTYPES = tuple(
+    dtype
+    for dtype in (
+        getattr(torch, "float8_e4m3fn", None),
+        getattr(torch, "float8_e5m2", None),
+    )
+    if dtype is not None
+)
+
+
 def get_lora_parameters(proj):
     """Return (weight, weight quant_state, lora A, lora B, lora scale).
     With QAT enabled, also fake-quantizes the base layer and lora weights.
@@ -298,9 +313,11 @@ def get_lora_parameters(proj):
         if weight_fake_quantizer is not None:
             W = weight_fake_quantizer(W)
 
-    # Get quant state for 4bit or FP8
+    # Get quant state for 4bit or FP8. Only fall back to a weight_scale(_inv) when the
+    # weight is still fp8; a bf16 weight (e.g. a decompressed compressed-tensors layer)
+    # must not carry a scale as its quant state or fast_gemv will crash on it.
     W_quant = getattr(W, "quant_state", None)
-    if W_quant is None:
+    if W_quant is None and W.dtype in _FP8_WEIGHT_DTYPES:
         W_quant = getattr(base_layer, "weight_scale_inv", None)
         if W_quant is None:
             W_quant = getattr(base_layer, "weight_scale", None)
@@ -349,9 +366,11 @@ def get_lora_parameters_bias(proj):
     )  # (proj.base_layer if hasattr(proj, "base_layer") else proj)
     W = base_layer.weight
 
-    # Get quant state for 4bit or FP8
+    # Get quant state for 4bit or FP8. Only fall back to a weight_scale(_inv) when the
+    # weight is still fp8; a bf16 weight (e.g. a decompressed compressed-tensors layer)
+    # must not carry a scale as its quant state or fast_gemv will crash on it.
     W_quant = getattr(W, "quant_state", None)
-    if W_quant is None:
+    if W_quant is None and W.dtype in _FP8_WEIGHT_DTYPES:
         W_quant = getattr(base_layer, "weight_scale_inv", None)
         if W_quant is None:
             W_quant = getattr(base_layer, "weight_scale", None)
diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index 1c75f8ce66..ae3b94ecc8 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-__version__ = "2026.7.1"
+__version__ = "2026.7.2"
 
 __all__ = [
     "SUPPORTS_BFLOAT16",
@@ -86,6 +86,8 @@ __all__ = [
     "maybe_prefetch_hf_snapshot",
     "is_moe_model",
     "get_moe_target_parameters",
+    "get_moe_target_modules",
+    "warn_if_zoo_cannot_merge_moe_experts",
     "_select_moe_detection_targets",
     "make_fast_generate_wrapper",
     "_mark_unsloth_disable_data_parallel",
@@ -422,7 +424,7 @@ def apply_unsloth_gradient_checkpointing(use_gradient_checkpointing, max_seq_len
 # access on some GPU architectures (B200). Falls back to eager safely.
 _FLEX_EXCLUDED_MODELS = ("gpt_oss", "mllama", "nemotron_h", "modernbert")
 _FLEX_PREFERRED_MODELS = ("gemma3", "gemma3_text", "shieldgemma2")
-_SDPA_EXCLUDED_MODELS = ("gpt_oss",)
+_SDPA_EXCLUDED_MODELS = ("gpt_oss", "deepseek_v4")
 # The loader (loader.py) forces supports_sdpa=False for these because their bundled
 # SDPA modules are wrong. Kept here, not in loader.py, so _is_sdpa_excluded can honor
 # them without a loader -> _utils import cycle (loader.py already imports from _utils
@@ -435,8 +437,10 @@ DISABLE_SDPA_MODEL_NAMES = [
     "gemma3_text",  # Gemma3TextModel (EmbeddingGemma) - substring match, keep underscore
     "gpt_oss",
 ]
-_FLASH_EXCLUDED_MODELS = ("gpt_oss",)
-_EAGER_ONLY_PREFIXES = ("gemma3n",)
+_FLASH_EXCLUDED_MODELS = ("gpt_oss", "deepseek_v4")
+# deepseek_v4's custom attention is sdpa/flash-incompatible; force eager, and
+# excluded above so an explicit sdpa/flash request cannot re-enable the crash.
+_EAGER_ONLY_PREFIXES = ("gemma3n", "deepseek_v4")
 _FLASH_ATTENTION_MAX_HEAD_DIM = 256
 _FLASH_ATTENTION_DISABLED_WARNED = set()
 
@@ -4060,13 +4064,17 @@ def get_moe_target_parameters(model, target_modules = None) -> Optional[List[str
         alternate_name = "experts.down_proj",
     )
 
-    # gate_up_proj combines both gate_proj and up_proj in MoE
-    # Also match "gate_up_proj" directly since users may specify the fused name
+    # gate_up_proj combines gate_proj and up_proj; also match the fused name directly.
+    # Only target a fused expert Parameter that exists: per-expert Linear layouts
+    # (e.g. gpt-oss bnb-4bit) have no fused Parameter and are handled by
+    # get_moe_target_modules, so skip them rather than pass PEFT a dead path.
     if "gate_proj" in target_set or "up_proj" in target_set or "gate_up_proj" in target_set:
-        moe_params.append(gate_up_name)
+        if _moe_parameter_exists(model, gate_up_name):
+            moe_params.append(gate_up_name)
 
     if "down_proj" in target_set:
-        moe_params.append(down_name)
+        if _moe_parameter_exists(model, down_name):
+            moe_params.append(down_name)
 
     if moe_params:
         print(
@@ -4077,6 +4085,111 @@ def get_moe_target_parameters(model, target_modules = None) -> Optional[List[str
     return None
 
 
+def _moe_parameter_exists(model, name: str) -> bool:
+    """True if ``name`` is an exact suffix of some parameter path on the model."""
+    if not hasattr(model, "named_parameters"):
+        return False
+    try:
+        for parameter_name, _ in model.named_parameters():
+            if parameter_name == name or parameter_name.endswith("." + name):
+                return True
+    except Exception:
+        return False
+    return False
+
+
+def get_moe_target_modules(model, target_modules = None) -> List[str]:
+    """Per-expert ``target_modules`` suffixes for MoE models whose experts are stored
+    as per-expert ``nn.Linear`` ModuleLists rather than fused nn.Parameters.
+
+    gpt-oss bnb-4bit is the canonical case (mlp.experts.gate_up_projs.<i> /
+    down_projs.<i> as Linear4bit): no fused Parameter, and the plain
+    gate/up/down_proj leaves do not match, so LoRA skips them. Returning the
+    per-expert suffixes makes PEFT attach via ordinary suffix matching (the
+    module-LoRA counterpart of get_moe_target_parameters). Returns [] for non-MoE,
+    fused-parameter MoEs, an absent per-expert layout, or a request that omits the
+    MLP experts (so an attention-only run does not train experts).
+    """
+    if not is_moe_model(model):
+        return []
+    if target_modules is None:
+        return []
+    if isinstance(target_modules, str):
+        target_set = _moe_target_set_from_string(target_modules)
+    else:
+        target_set = {
+            target
+            for target in target_modules or ()
+            if (isinstance(target, str) and "." not in target and target in _MOE_BROAD_MLP_TARGETS)
+        }
+    if not (target_set & _MOE_BROAD_MLP_TARGETS):
+        return []
+
+    if not hasattr(model, "named_modules"):
+        return []
+
+    # Scope the returned suffixes to the requested projection leaves, matching
+    # get_moe_target_parameters: gate_proj/up_proj/gate_up_proj map to the fused
+    # gate_up ModuleList (e.g. gate_up_projs); down_proj maps to the down ModuleList
+    # (e.g. down_projs). A down-only (or gate/up-only) request must not pull in the
+    # other projection.
+    want_gate_up = bool(target_set & {"gate_proj", "up_proj", "gate_up_proj"})
+    want_down = "down_proj" in target_set
+
+    targets = set()
+    for name, module in model.named_modules():
+        if not isinstance(module, torch.nn.ModuleList) or len(module) == 0:
+            continue
+        parent, _, leaf = name.rpartition(".")
+        # ModuleList directly under an ``experts`` container, holding only Linear
+        # leaves (bnb Linear4bit / Linear8bitLt subclass nn.Linear). After PEFT has
+        # wrapped the experts the child is a LoRA layer whose ``base_layer`` is the
+        # Linear, so accept that too (keeps this idempotent across a re-wrapped model).
+        if not parent.endswith("experts"):
+            continue
+        if not all(
+            isinstance(child, torch.nn.Linear)
+            or isinstance(getattr(child, "base_layer", None), torch.nn.Linear)
+            for child in module
+        ):
+            continue
+        # Honor the requested subset: classify the ModuleList by projection role.
+        leaf_lower = leaf.lower()
+        is_down = "down" in leaf_lower
+        is_gate_up = (not is_down) and ("gate" in leaf_lower or "up" in leaf_lower)
+        if is_down and not want_down:
+            continue
+        if is_gate_up and not want_gate_up:
+            continue
+        # One entry per expert index; ``leaf.<i>`` matches expert i in every layer.
+        for expert_index in range(len(module)):
+            targets.add(f"{leaf}.{expert_index}")
+
+    return sorted(targets)
+
+
+def warn_if_zoo_cannot_merge_moe_experts():
+    """Warn once when the installed unsloth_zoo cannot fold per-expert Linear MoE LoRA
+    into a merged_16bit checkpoint. Older zoo releases keep the fused gate_up_proj /
+    down_proj tensors and drop the per-expert gate_up_projs.<i> / down_projs.<i> deltas,
+    so save_pretrained_merged("merged_16bit") would silently lose the expert training
+    (the LoRA adapter itself still saves and reloads correctly)."""
+    try:
+        from unsloth_zoo import saving_utils as _saving_utils
+
+        # _fold_perexpert_lora_into_fused is the helper that folds these experts.
+        if hasattr(_saving_utils, "_fold_perexpert_lora_into_fused"):
+            return
+    except Exception:
+        return  # cannot introspect zoo -> stay quiet rather than false-alarm
+    logger.warning_once(
+        "Unsloth: the installed unsloth_zoo will not fold these per-expert experts into "
+        "a merged_16bit checkpoint, so save_pretrained_merged('merged_16bit') would drop "
+        "the expert LoRA. Upgrade unsloth_zoo to merge them; saving the LoRA adapter is "
+        "unaffected."
+    )
+
+
 def _select_moe_detection_targets(
     original_target_modules,
     scoped_target_modules,
diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index a1da099758..defacadfee 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -28,7 +28,11 @@ from ._utils import (
     is_bfloat16_supported,
     get_quant_type,
 )
-from .loader_utils import _exclude_rope_inv_freq_from_ddp, _get_fp8_mode_and_check_settings
+from .loader_utils import (
+    _exclude_rope_inv_freq_from_ddp,
+    _get_fp8_mode_and_check_settings,
+    _restore_dropped_fp8_scales,
+)
 from ..utils.packing import (
     get_packed_info_from_kwargs,
     mask_packed_sequence_boundaries,
@@ -1651,6 +1655,26 @@ def _rope_scaling_as_dict(rope_scaling):
         return {}
 
 
+def _extended_rope_scaling(config, factor):
+    """RoPE scaling to extend a model past its native window. Keeps native llama3 as-is
+    (linear extension is far worse for long context); everything else gets linear. Returns
+    (scaling_or_None, type): None keeps llama3. The linear dict carries rope_theta so
+    transformers v5 (which stores it under rope_parameters) keeps the real base, not 10000.
+    Only llama3 is preserved because patch_llama_rope_scaling can only rebuild linear/llama3/
+    longrope and its longrope branch needs a top-level original_max_position_embeddings."""
+    existing = _rope_scaling_as_dict(
+        getattr(config, "rope_scaling", None) or getattr(config, "rope_parameters", None) or {}
+    )
+    existing_type = existing.get("rope_type") or existing.get("type")
+    if existing_type == "llama3":
+        return None, existing_type
+    return {
+        "type": "linear",
+        "factor": factor,
+        "rope_theta": _get_rope_theta(config),
+    }, existing_type
+
+
 def _llama3_inv_freq_from_config(
     config,
     rope_scaling,
@@ -2518,34 +2542,33 @@ class FastLlamaModel:
             max_seq_length = model_max_seq_length
 
         if (rope_scaling is None) and (max_seq_length > model_max_seq_length):
-            rope_scaling = max_seq_length / model_max_seq_length
+            factor = max_seq_length / model_max_seq_length
 
             if fast_inference:
                 raise NotImplementedError(
                     "Unsloth: Fast inference does not yet work with RoPE Scaling."
                 )
 
-            logger.warning_once(
-                f"Unsloth: {model_name} can only handle sequence lengths of at most "
-                f"{model_max_seq_length}.\nBut with kaiokendev's RoPE scaling of "
-                f"{round(rope_scaling, 3)}, it can be magically be extended to "
-                f"{max_seq_length}!"
-            )
-
-            # Warn RoPE scaling isn't allowed
-            if not has_rope_scaling:
-                raise RuntimeError(
-                    f"However, {model_name} doesn't support RoPE Scaling!\n"
-                    "Please file a feature request at https://github.com/unslothai/unsloth."
+            linear_scaling, native_type = _extended_rope_scaling(model_config, factor)
+            if linear_scaling is not None:
+                logger.warning_once(
+                    f"Unsloth: {model_name} can only handle sequence lengths of at most "
+                    f"{model_max_seq_length}.\nBut with kaiokendev's RoPE scaling of "
+                    f"{round(factor, 3)}, it can be magically be extended to "
+                    f"{max_seq_length}!"
+                )
+                if not has_rope_scaling:
+                    raise RuntimeError(
+                        f"However, {model_name} doesn't support RoPE Scaling!\n"
+                        "Please file a feature request at https://github.com/unslothai/unsloth."
+                    )
+                kwargs["rope_scaling"] = linear_scaling
+            else:
+                # Native llama3 scaling already handles long context; just widen the window.
+                logger.warning_once(
+                    f"Unsloth: extending {model_name} to {max_seq_length} using its native "
+                    f"{native_type} RoPE scaling."
                 )
-
-            rope_scaling = {
-                "type": "linear",
-                "factor": rope_scaling,
-            }
-
-            # Add to kwargs
-            kwargs["rope_scaling"] = rope_scaling
 
         from .loader_utils import (
             check_and_disable_bitsandbytes_loading,
@@ -2659,6 +2682,18 @@ class FastLlamaModel:
                     offload_embedding = False,
                     fast_inference = fast_inference,
                 )
+                # Re-apply block-fp8 weight_scale_inv tensors transformers dropped on load (#6200).
+                _restore_dropped_fp8_scales(
+                    model,
+                    model_name,
+                    local_files_only = kwargs.get("local_files_only", False),
+                    token = token,
+                    # Weights load from the default branch (revision not forwarded), so read scales from there too.
+                    revision = None,
+                    subfolder = kwargs.get("subfolder"),
+                    cache_dir = kwargs.get("cache_dir"),
+                    variant = kwargs.get("variant"),
+                )
             elif not fast_inference:
                 if user_config is not None:
                     # Transformers 5.x @strict model init rejects extra kwargs next
@@ -2697,6 +2732,18 @@ class FastLlamaModel:
                     offload_embedding = False,
                     fast_inference = False,
                 )
+                # Re-apply block-fp8 weight_scale_inv tensors transformers dropped on load (#6200).
+                _restore_dropped_fp8_scales(
+                    model,
+                    model_name,
+                    local_files_only = kwargs.get("local_files_only", False),
+                    token = token,
+                    # Weights load from the default branch (revision not forwarded), so read scales from there too.
+                    revision = None,
+                    subfolder = kwargs.get("subfolder"),
+                    cache_dir = kwargs.get("cache_dir"),
+                    variant = kwargs.get("variant"),
+                )
                 model.fast_generate = make_fast_generate_wrapper(model.generate)
                 model.fast_generate_batches = None
             else:
@@ -3105,6 +3152,11 @@ class FastLlamaModel:
             new_target_modules = list(target_modules) + list(
                 modules_to_save if modules_to_save is not None else []
             )
+            # Per-expert Linear MoE experts (e.g. gpt-oss bnb-4bit) were auto-added to the
+            # saved target_modules when the adapter was first created. Recompute them so a
+            # repeat get_peft_model call with the same args stays idempotent instead of
+            # tripping the mismatch below. No-op for non per-expert-Linear models.
+            new_target_modules += get_moe_target_modules(model, target_modules)
 
             # Now check!
             new_target_modules = set(new_target_modules)
@@ -3331,6 +3383,19 @@ class FastLlamaModel:
         if target_parameters is None:
             target_parameters = get_moe_target_parameters(model, target_modules)
 
+        # Per-expert Linear expert layouts (e.g. gpt-oss bnb-4bit) are Linear modules,
+        # not fused Parameters, so target them via target_modules. No-op otherwise.
+        _moe_module_targets = get_moe_target_modules(model, target_modules)
+        if _moe_module_targets:
+            _added = [t for t in _moe_module_targets if t not in final_modules]
+            final_modules.extend(_added)
+            if _added:
+                print(
+                    f"Unsloth: Detected MoE model with per-expert Linear experts. "
+                    f"Enabling LoRA on {len(_added)} expert projection modules."
+                )
+                warn_if_zoo_cannot_merge_moe_experts()
+
         if finetune_last_n_layers is not None and layers_to_transform is None:
             from .vision import _get_total_transformer_layers
             _total_layers = _get_total_transformer_layers(model)
diff --git a/unsloth/models/loader_utils.py b/unsloth/models/loader_utils.py
index d6d6ce877e..1e9cc641e9 100644
--- a/unsloth/models/loader_utils.py
+++ b/unsloth/models/loader_utils.py
@@ -239,6 +239,11 @@ def get_model_name(
         and new_model_name.lower() in BAD_MAPPINGS
     ):
         new_model_name = BAD_MAPPINGS[new_model_name.lower()]
+    elif new_model_name is None and model_name.lower() in BAD_MAPPINGS:
+        # Some bad names (e.g. the `-unsloth-bnb-4bit` dynamic quants) are keys
+        # of the mappers, not values, so the resolver returns None for them and
+        # the remap above is skipped; remap the input name directly instead.
+        new_model_name = BAD_MAPPINGS[model_name.lower()]
 
     if (
         new_model_name is None
@@ -362,6 +367,287 @@ def _tag_model_with_fp8_torchao_config(model: torch.nn.Module, fp8_mode: str):
         pass
 
 
+_FP8_DTYPES = tuple(
+    dtype
+    for dtype in (getattr(torch, "float8_e4m3fn", None), getattr(torch, "float8_e5m2", None))
+    if dtype is not None
+)
+
+
+def _fp8_block_size_from_config(model):
+    """Return the [block_out, block_in] block size of an fp8 checkpoint, or None if not block-fp8."""
+    config = getattr(model, "config", None)
+    quant = getattr(config, "quantization_config", None)
+    if quant is None:
+        return None
+    if hasattr(quant, "to_dict"):
+        quant = quant.to_dict()
+    if not isinstance(quant, dict):
+        return None
+    if quant.get("quant_method") != "fp8":
+        return None
+    block = quant.get("weight_block_size")
+    if not block:
+        return None
+    if isinstance(block, (int, float)):
+        block = [block, block]
+    elif isinstance(block, (list, tuple)):
+        if len(block) == 1:
+            block = [block[0], block[0]]
+        elif len(block) < 2:
+            return None
+    else:
+        return None
+    return [int(block[0]), int(block[1])]
+
+
+def _load_fp8_weight_map(
+    model_name,
+    local_files_only,
+    token,
+    revision = None,
+    subfolder = None,
+    cache_dir = None,
+):
+    """Return the checkpoint's tensor->file map, using the same snapshot the load used.
+
+    Prefers the sharded `model.safetensors.index.json`; falls back to a single `model.safetensors`
+    (every tensor maps to that one file) so unsharded checkpoints are covered too.
+    """
+
+    def _local_path(filename):
+        return (
+            os.path.join(model_name, subfolder, filename)
+            if subfolder
+            else os.path.join(model_name, filename)
+        )
+
+    def _remote_path(filename):
+        from huggingface_hub import hf_hub_download
+        return hf_hub_download(
+            model_name,
+            filename,
+            revision = revision,
+            subfolder = subfolder,
+            cache_dir = cache_dir,
+            local_files_only = local_files_only,
+            token = token,
+        )
+
+    index_file = "model.safetensors.index.json"
+    single_file = "model.safetensors"
+    is_local = os.path.isdir(model_name)
+
+    # Sharded checkpoint.
+    if is_local and os.path.exists(_local_path(index_file)):
+        index_path = _local_path(index_file)
+    elif not is_local:
+        try:
+            index_path = _remote_path(index_file)
+        except Exception:
+            index_path = None
+    else:
+        index_path = None
+    if index_path is not None:
+        import json
+        with open(index_path, "r") as f:
+            return json.load(f).get("weight_map", None)
+
+    # Unsharded single file: map every tensor to it.
+    try:
+        if is_local and os.path.exists(_local_path(single_file)):
+            single_path = _local_path(single_file)
+        elif not is_local:
+            single_path = _remote_path(single_file)
+        else:
+            return None
+        from safetensors import safe_open
+        with safe_open(single_path, framework = "pt") as f:
+            return {key: single_file for key in f.keys()}
+    except Exception:
+        return None
+
+
+def _resolve_fp8_shard(
+    model_name,
+    shard,
+    local_files_only,
+    token,
+    revision = None,
+    subfolder = None,
+    cache_dir = None,
+):
+    """Resolve a checkpoint shard filename to a local path (repo id or local dir)."""
+    if os.path.isdir(model_name):
+        return (
+            os.path.join(model_name, subfolder, shard)
+            if subfolder
+            else os.path.join(model_name, shard)
+        )
+    from huggingface_hub import hf_hub_download
+
+    return hf_hub_download(
+        model_name,
+        shard,
+        revision = revision,
+        subfolder = subfolder,
+        cache_dir = cache_dir,
+        local_files_only = local_files_only,
+        token = token,
+    )
+
+
+def _match_fp8_module(module_by_name, base):
+    """Resolve a checkpoint module name to a live module, allowing for VLM key remappings.
+
+    VLM loads can name the text tower differently from the checkpoint keys: `text_only=True`
+    strips the `language_model.` wrapper (so `model.language_model.layers.*` -> `model.layers.*`),
+    and full VLM loads may expose `model.language_model.*` while the checkpoint stores
+    `language_model.model.*`. Try the raw key first, then a few safe remappings.
+    """
+    if base in module_by_name:
+        return module_by_name[base]
+    candidates = []
+    if "language_model." in base:
+        candidates.append(base.replace("language_model.", "", 1))  # text-only: drop wrapper
+    if "language_model.model." in base:
+        candidates.append(base.replace("language_model.model.", "model.language_model.", 1))
+    if base.startswith("language_model."):
+        candidates.append("model." + base)  # add model. prefix
+    for candidate in candidates:
+        if candidate in module_by_name:
+            return module_by_name[candidate]
+    return None
+
+
+def _restore_dropped_fp8_scales(
+    model,
+    model_name,
+    *,
+    local_files_only = False,
+    token = None,
+    revision = None,
+    subfolder = None,
+    cache_dir = None,
+    variant = None,
+):
+    """Re-apply block-fp8 `weight_scale_inv` tensors that transformers dropped on load.
+
+    On some block-scale fp8 checkpoints (e.g. Qwen3.6-27B-FP8, issue #6200) transformers fails to
+    convert a Linear (such as `mlp.gate_proj`) to an fp8 module, loading the raw quantized values
+    into a plain bf16 weight and discarding its `weight_scale_inv` as an unexpected key. The weight
+    is then used un-scaled, producing a garbage model. For every checkpoint scale whose live weight
+    is not fp8, dequantize the orphaned weight in place. Modules that were converted correctly keep
+    an fp8 weight and are skipped, so a healthy checkpoint is a no-op. Returns (restored, skipped).
+    """
+    try:
+        block = _fp8_block_size_from_config(model)
+        if block is None or not _FP8_DTYPES:
+            return (0, 0)
+        # A variant load reads variant-named files; skip to avoid applying default scales to them.
+        if variant:
+            return (0, 0)
+        # No fp8 params means the checkpoint was dequantized on purpose (e.g. load_in_16bit);
+        # re-applying a scale would corrupt those already-correct 16bit weights, so do nothing.
+        if not any(p.dtype in _FP8_DTYPES for p in model.parameters()):
+            return (0, 0)
+        weight_map = _load_fp8_weight_map(
+            model_name, local_files_only, token, revision, subfolder, cache_dir
+        )
+        if not weight_map:
+            return (0, 0)
+
+        scale_keys = {k: v for k, v in weight_map.items() if k.endswith(".weight_scale_inv")}
+        if not scale_keys:
+            return (0, 0)
+
+        module_by_name = dict(model.named_modules())
+        bs0, bs1 = block
+        restored = 0
+        skipped = 0
+        failed = 0
+        offloaded = 0
+        shard_cache = {}
+        for scale_key, shard in scale_keys.items():
+            base = scale_key[: -len(".weight_scale_inv")]
+            module = _match_fp8_module(module_by_name, base)
+            if module is None:
+                continue
+            weight = getattr(module, "weight", None)
+            if not isinstance(weight, torch.Tensor) or weight.ndim != 2:
+                continue
+            if weight.device.type == "meta":
+                # Disk-offloaded layer: weight lives on meta until forward, so it cannot be
+                # scaled in place here. Count and warn rather than silently leave it unscaled.
+                offloaded += 1
+                continue
+            if weight.dtype in _FP8_DTYPES:
+                # Correctly converted fp8 module: the fp8 path already handles the scale.
+                skipped += 1
+                continue
+
+            # Errors after this point are per-tensor: warn and continue, never abort or hide them.
+            try:
+                if shard not in shard_cache:
+                    from safetensors import safe_open
+                    shard_path = _resolve_fp8_shard(
+                        model_name,
+                        shard,
+                        local_files_only,
+                        token,
+                        revision,
+                        subfolder,
+                        cache_dir,
+                    )
+                    shard_cache[shard] = safe_open(shard_path, framework = "pt")
+                scale = shard_cache[shard].get_tensor(scale_key).to(torch.float32)
+
+                out_features, in_features = weight.shape
+                out_blocks = (out_features + bs0 - 1) // bs0
+                in_blocks = (in_features + bs1 - 1) // bs1
+                if tuple(scale.shape) == (out_blocks, in_blocks):
+                    pass
+                elif tuple(scale.shape) == (in_blocks, out_blocks) and out_blocks != in_blocks:
+                    # Transposed block layout: same handling as the fp8 forward path.
+                    scale = scale.t().contiguous()
+                else:
+                    # Shape does not match the block grid: skip rather than apply a wrong scale.
+                    continue
+                scale = scale.to(weight.device)
+                with torch.no_grad():
+                    if out_features % bs0 == 0 and in_features % bs1 == 0:
+                        # Memory-frugal path: multiply block views in place against the broadcast
+                        # fp32 scale, avoiding a full expanded scale and fp32 copy that could OOM.
+                        # The in-place multiply promotes to fp32, matching the fallback exactly.
+                        module.weight.data.view(out_blocks, bs0, in_blocks, bs1).mul_(
+                            scale[:, None, :, None]
+                        )
+                    else:
+                        scale_expanded = scale.repeat_interleave(bs0, dim = 0).repeat_interleave(
+                            bs1, dim = 1
+                        )[:out_features, :in_features]
+                        module.weight.data = (weight.to(torch.float32) * scale_expanded).to(
+                            weight.dtype
+                        )
+                restored += 1
+            except Exception:
+                failed += 1
+                continue
+
+        if restored > 0:
+            print(f"Unsloth: Restored {restored} dropped FP8 weight_scale_inv tensor(s) on load")
+        if failed > 0:
+            print(f"Unsloth: {failed} dropped FP8 weight_scale_inv tensor(s) could not be restored")
+        if offloaded > 0:
+            print(
+                f"Unsloth: {offloaded} dropped FP8 weight_scale_inv tensor(s) skipped because the "
+                "layer is disk-offloaded; load without disk offload so the scales can be restored"
+            )
+        return (restored, skipped)
+    except Exception:
+        return (0, 0)
+
+
 def check_and_disable_bitsandbytes_loading(
     model_config,
     load_in_4bit = True,
diff --git a/unsloth/models/mapper.py b/unsloth/models/mapper.py
index 57c1e292c3..f3a0e1f9bb 100644
--- a/unsloth/models/mapper.py
+++ b/unsloth/models/mapper.py
@@ -134,7 +134,7 @@ __INT_TO_FLOAT_MAPPER = \
         "unsloth/gemma-7b-it",
         "google/gemma-7b-it",
     ),
-    "unsloth/gemma-2b-bnb-4bit" : (
+    "unsloth/gemma-2b-it-bnb-4bit" : (
         "unsloth/gemma-2b-it",
         "google/gemma-2b-it",
     ),
diff --git a/unsloth/models/rl.py b/unsloth/models/rl.py
index b5cadf2dea..eeab8fbaca 100644
--- a/unsloth/models/rl.py
+++ b/unsloth/models/rl.py
@@ -1651,7 +1651,36 @@ def _patch_trl_rl_trainers_impl(trainer_file = "grpo_trainer"):
 
         RLTrainer_source = re.sub(pattern, new_options, RLTrainer_source, flags = re.DOTALL)
 
-        if trl_version >= Version("0.27.0"):
+        if trl_version >= Version("1.4.0"):
+            # The `elif is_peft_model(model) and args.beta != 0.0:` ref-adapter block
+            # was introduced in TRL 1.4.0 and is used through 1.7.x. Remove only that
+            # block, anchored on the final ref_param copy so we do NOT also swallow the
+            # following gradient-checkpointing enable_input_require_grads() block.
+            peft_pattern = (
+                r"\s*elif is_peft_model\(model\) and args\.beta != 0\.0:"
+                r".*?"
+                r"ref_param\.data\.copy_\(param\.data\)"
+            )
+
+            replacement_comment = (
+                "\n        # PEFT initialization logic removed via script for trl >= 1.4.0\n"
+            )
+
+            RLTrainer_source = re.sub(
+                peft_pattern, replacement_comment, RLTrainer_source, flags = re.DOTALL
+            )
+
+            if trl_version >= Version("1.7.0"):
+                # router_aux_loss_coef / aux_loss_enabled were added in TRL 1.7.0. Unsloth's
+                # optimized GRPO forward cannot compute the MoE router aux loss, so reject
+                # explicit opt-in (router_aux_loss_coef > 0) at init rather than silently ignoring it.
+                RLTrainer_source = RLTrainer_source.replace(
+                    "self.aux_loss_enabled = is_moe and args.router_aux_loss_coef != 0.0",
+                    "self.aux_loss_enabled = is_moe and args.router_aux_loss_coef != 0.0\n"
+                    '        if self.aux_loss_enabled: raise NotImplementedError("Unsloth GRPO does not compute the MoE router auxiliary loss; set router_aux_loss_coef = 0 (the Unsloth default).")',
+                )
+
+        elif trl_version >= Version("0.27.0"):
             peft_pattern = (
                 r"\s*if is_peft_available\(\) and is_peft_model\(model\) and args\.beta != 0\.0:"
                 r".*?"
@@ -1689,6 +1718,11 @@ def _patch_trl_rl_trainers_impl(trainer_file = "grpo_trainer"):
         'if getattr(model, "is_loaded_in_4bit", False) or getattr(model, "is_loaded_in_8bit", False):',
         "if False:",
     )
+    # TRL >= 1.7.0 spells the same QLoRA bf16 cast as `if _is_quantized_model:`.
+    RLTrainer_source = RLTrainer_source.replace(
+        "if _is_quantized_model:",
+        "if False:",
+    )
 
     if RLTrainer_name == "SFTTrainer":
         original_text = (
diff --git a/unsloth/models/rl_replacements.py b/unsloth/models/rl_replacements.py
index 098950de08..ffb845b04f 100644
--- a/unsloth/models/rl_replacements.py
+++ b/unsloth/models/rl_replacements.py
@@ -1203,6 +1203,8 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
                 if os.environ.get("UNSLOTH_FORCE_FLOAT32", "0") == "1":
                     self._autocast_dtype = torch.float16
 
+            compute_aux_loss = kwargs.get("compute_aux_loss", None)
+
             pixel_values, image_grid_thw = (
                 kwargs.get("pixel_values", None),
                 kwargs.get("image_grid_thw", None),
@@ -1846,7 +1848,7 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
                         _extra_vision_kwargs["mm_token_type_ids"] = mm_token_type_ids_chunk
                     with torch.amp.autocast(device_type = "cuda", dtype = self._autocast_dtype):
                         if pixel_values is None:
-                            logits_chunk = unwrapped_model(
+                            outputs = unwrapped_model(
                                 input_ids = input_ids_chunk,
                                 attention_mask = attention_mask_chunk,
                                 pixel_values = pixel_values_chunk,
@@ -1854,7 +1856,10 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
                                 pixel_attention_mask = pixel_attention_mask_chunk,
                                 image_sizes = image_sizes_chunk,
                                 **_extra_vision_kwargs,
-                            ).logits
+                            )
+
+                            logits_chunk = outputs.logits
+                            del outputs  # free hidden_states before chunked log-softmax
 
                             completion_input_ids_chunk = input_ids_chunk[
                                 :, -(logits_to_keep + max_left_pad) :
@@ -1876,7 +1881,7 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
                         else:
                             # Essentially, for VLMs we do not go via the optimized path in models/,
                             # so we don't encounter the Flash Attn left-padding issue.
-                            logits_chunk = unwrapped_model(
+                            outputs = unwrapped_model(
                                 input_ids = input_ids_chunk,
                                 attention_mask = attention_mask_chunk,
                                 pixel_values = pixel_values_chunk,
@@ -1885,7 +1890,10 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
                                 image_sizes = image_sizes_chunk,
                                 logits_to_keep = logits_to_keep + 1,
                                 **_extra_vision_kwargs,
-                            ).logits
+                            )
+
+                            logits_chunk = outputs.logits
+                            del outputs  # free hidden_states before chunked log-softmax
 
                             logits_chunk = logits_chunk[:, :-1, :]
                             completion_input_ids_chunk = input_ids_chunk[:, -logits_to_keep:]
@@ -1914,11 +1922,15 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
                     all_logprobs_list.append(logprobs_chunk)
                 if logprobs is None:  # padded fallback when packing was not used
                     logprobs = torch.cat(all_logprobs_list, dim = 0)
+
                 entropies = None
 
             os.environ["UNSLOTH_RETURN_HIDDEN_STATES"] = "0"
-
-            return logprobs.detach(), entropies  # logps, entropies
+            # aux loss is unused: it is off by default (router_aux_loss_coef set to 0 in models/rl.py)
+            # and explicit opt-in is rejected at trainer init, so this is always None (kept in the
+            # return for TRL >= 1.7.0's 3-tuple contract).
+            aux_loss = None
+            return logprobs.detach(), entropies, aux_loss  # logps, entropies, aux_loss
             # input_ids = input_ids[:, -logits_to_keep:]
             # For transformers<=4.48, logits_to_keep argument isn't supported, so here we drop logits ourselves.
             # See https://github.com/huggingface/trl/issues/2770
@@ -1937,6 +1949,24 @@ def grpo_trainer__get_per_token_logps_and_entropies(function_name, function):
             # return  logps #  compute logprobs for the input tokens
 
     function = inspect.getsource(_get_per_token_logps_and_entropies)
+    if trl_version < Version("1.7.0"):
+        # TRL < 1.7.0 unpacks (logps, entropies) at every call site; TRL >= 1.7.0
+        # always unpacks (logps, entropies, aux_loss). Drop the aux_loss element so
+        # the return arity matches the installed TRL. Regex tolerates comment /
+        # whitespace drift on the return line; fail loud if the anchor ever stops
+        # matching rather than silently shipping a 3-tuple to older TRL.
+        new_function, n = re.subn(
+            r"return (logprobs\.detach\(\), entropies), aux_loss[^\n]*",
+            r"return \1  # logps, entropies",
+            function,
+        )
+        if n != 1:
+            raise RuntimeError(
+                "Unsloth GRPO: could not downgrade the per-token-logps return to a "
+                f"2-tuple for TRL {trl_version} (matched {n} times, expected 1). The "
+                "return line changed; update the arity gate in rl_replacements.py."
+            )
+        function = new_function
     return function
 
 
diff --git a/unsloth/models/sentence_transformer.py b/unsloth/models/sentence_transformer.py
index c1172faa94..4a1a555bcf 100644
--- a/unsloth/models/sentence_transformer.py
+++ b/unsloth/models/sentence_transformer.py
@@ -990,7 +990,17 @@ class FastSentenceTransformer(FastModel):
             return None
 
     @staticmethod
-    def _create_transformer_module(model_name, model, tokenizer, max_seq_length, trust_remote_code):
+    def _create_transformer_module(
+        model_name,
+        model,
+        tokenizer,
+        max_seq_length,
+        trust_remote_code,
+        token = None,
+        cache_dir = None,
+        revision = None,
+        module_subfolder = "",
+    ):
         """Helper to create and configure a Transformer module."""
         from sentence_transformers.models import Transformer
 
@@ -1077,7 +1087,45 @@ class FastSentenceTransformer(FastModel):
                 elif "tokenizer_args" in transformer_init_params:
                     transformer_kwargs["tokenizer_args"] = trust_remote_code_kwargs.copy()
 
-                transformer_module = Transformer(model_name, **transformer_kwargs)
+                # Build via Transformer.load so the saved modality_config is honored: plain
+                # Transformer(...) makes ST 5.x infer a "message" modality for chat-template
+                # models (e.g. Qwen3-Embedding), chat-wrapping inputs and degrading embeddings
+                # (#6881). Only use .load when it resolves a Hub id (accepts the kwargs or
+                # **kwargs); legacy ST 3.x/4.x load(input_path) is local-only with no modality
+                # bug, so fall back to the constructor.
+                transformer_module = None
+                transformer_load = getattr(Transformer, "load", None)
+                has_modules_json = (
+                    FastSentenceTransformer._module_path(
+                        model_name, token, cache_dir = cache_dir, revision = revision
+                    )
+                    is not None
+                )
+                if callable(transformer_load) and has_modules_json:
+                    load_params = inspect.signature(transformer_load).parameters
+                    accepts_var_kw = any(
+                        p.kind is inspect.Parameter.VAR_KEYWORD for p in load_params.values()
+                    )
+                    hub_capable = accepts_var_kw or any(
+                        key in load_params for key in ("token", "cache_folder", "revision")
+                    )
+                    if hub_capable:
+                        load_kwargs = {
+                            "token": token,
+                            "cache_folder": cache_dir,
+                            "revision": revision,
+                            "trust_remote_code": trust_remote_code,
+                            **transformer_kwargs,
+                        }
+                        # Resolve config/tokenizer from the module's saved subfolder
+                        # (modules.json "path"), like stock ST; "" (root) is a no-op.
+                        if module_subfolder:
+                            load_kwargs["subfolder"] = module_subfolder
+                        if not accepts_var_kw:
+                            load_kwargs = {k: v for k, v in load_kwargs.items() if k in load_params}
+                        transformer_module = Transformer.load(model_name, **load_kwargs)
+                if transformer_module is None:
+                    transformer_module = Transformer(model_name, **transformer_kwargs)
             finally:
                 # Restore original Auto* loading immediately
                 AutoModel.from_pretrained = original_model_from_pretrained
@@ -1191,6 +1239,10 @@ class FastSentenceTransformer(FastModel):
                         tokenizer,
                         max_seq_length,
                         trust_remote_code,
+                        token,
+                        cache_dir,
+                        revision,
+                        module_subfolder = module_config.get("path") or "",
                     )
                     modules[name] = transformer_module
                 else:
@@ -1226,7 +1278,14 @@ class FastSentenceTransformer(FastModel):
         )
 
         transformer_module = FastSentenceTransformer._create_transformer_module(
-            model_name, model, tokenizer, max_seq_length, trust_remote_code
+            model_name,
+            model,
+            tokenizer,
+            max_seq_length,
+            trust_remote_code,
+            token,
+            cache_dir,
+            revision,
         )
         modules["0"] = transformer_module
 
diff --git a/unsloth/models/vision.py b/unsloth/models/vision.py
index 0a68a49fee..3b5af39e69 100644
--- a/unsloth/models/vision.py
+++ b/unsloth/models/vision.py
@@ -41,7 +41,11 @@ from ._utils import (
     set_task_config_attr,
 )
 from ._utils import *
-from .loader_utils import _exclude_rope_inv_freq_from_ddp, _get_fp8_mode_and_check_settings
+from .loader_utils import (
+    _exclude_rope_inv_freq_from_ddp,
+    _get_fp8_mode_and_check_settings,
+    _restore_dropped_fp8_scales,
+)
 from ..save import patch_saving_functions
 from ..models.loader_utils import is_distributed
 from unsloth_zoo.gradient_checkpointing import (
@@ -1192,6 +1196,17 @@ class FastBaseModel:
                     offload_embedding = offload_embedding,
                     fast_inference = fast_inference,
                 )
+                # Re-apply block-fp8 weight_scale_inv tensors transformers dropped on load (#6200).
+                _restore_dropped_fp8_scales(
+                    model,
+                    model_name,
+                    local_files_only = local_files_only,
+                    token = token,
+                    revision = kwargs.get("revision"),
+                    subfolder = kwargs.get("subfolder"),
+                    cache_dir = kwargs.get("cache_dir"),
+                    variant = kwargs.get("variant"),
+                )
                 if hasattr(model, "generate"):
                     model.fast_generate = make_fast_generate_wrapper(model.generate)
                     model.fast_generate_batches = error_out_no_vllm
@@ -1807,6 +1822,35 @@ class FastBaseModel:
             )
             target_parameters = get_moe_target_parameters(model, _moe_targets)
 
+        # Per-expert Linear expert layouts (e.g. gpt-oss bnb-4bit) target experts via
+        # target_modules, not fused Parameters. Extend either form PEFT accepts: a leaf
+        # list (explicit) or a regex string (auto / all-linear / scoped). No-op otherwise.
+        _moe_module_detect = _select_moe_detection_targets(
+            _moe_detect_target,
+            target_modules,
+            finetune_mlp_modules = finetune_mlp_modules,
+            finetune_language_layers = finetune_language_layers,
+        )
+        _moe_module_targets = get_moe_target_modules(model, _moe_module_detect)
+        if _moe_module_targets:
+            if isinstance(target_modules, (list, tuple)):
+                target_modules = list(target_modules) + [
+                    target for target in _moe_module_targets if target not in target_modules
+                ]
+            elif isinstance(target_modules, str):
+                _expert_leaves = sorted({t.rsplit(".", 1)[0] for t in _moe_module_targets})
+                _expert_alt = (
+                    r".*\.experts\.(?:"
+                    + "|".join(re.escape(leaf) for leaf in _expert_leaves)
+                    + r")\.\d+"
+                )
+                target_modules = f"(?:{target_modules})|(?:{_expert_alt})"
+            print(
+                f"Unsloth: Detected MoE model with per-expert Linear experts. "
+                f"Enabling LoRA on {len(_moe_module_targets)} expert projection modules."
+            )
+            warn_if_zoo_cannot_merge_moe_experts()
+
         if finetune_last_n_layers is not None and layers_to_transform is None:
             _total_layers = _get_total_transformer_layers(model)
             if _total_layers is not None and _total_layers > 0:
diff --git a/unsloth_cli/_inference.py b/unsloth_cli/_inference.py
index e8a3414d73..c3b188710e 100644
--- a/unsloth_cli/_inference.py
+++ b/unsloth_cli/_inference.py
@@ -4,9 +4,11 @@
 """Model loading and streaming shared by `inference` and `chat`."""
 
 import asyncio
+import json
 import os
 import re
 import sys
+from contextlib import contextmanager, redirect_stderr, redirect_stdout
 from pathlib import Path
 from typing import List, Optional
 
@@ -14,10 +16,18 @@ import typer
 
 _THINK_OPEN = "<think>"
 _THINK_BLOCK = re.compile(rf"{re.escape(_THINK_OPEN)}.*?</think>", re.DOTALL)
+_STREAMED_ERROR_PREFIX = "Error: "
 
 # Cloudflare (in front of remote Studio proxies like RunPod) 403s the default
 # "Python-urllib/X.Y" User-Agent as a bot; send a real one on every request.
 _USER_AGENT = "unsloth-cli"
+_MPI_ENV_PAIRS = (
+    ("OMPI_COMM_WORLD_RANK", "OMPI_COMM_WORLD_SIZE"),
+    ("PMI_RANK", "PMI_SIZE"),
+    ("PMIX_RANK", "PMIX_SIZE"),
+    ("MPI_RANK", "MPI_WORLD_SIZE"),
+    ("MV2_COMM_WORLD_RANK", "MV2_COMM_WORLD_SIZE"),
+)
 
 # Built lazily; urllib stays function-local to match this module.
 _no_redirect_opener = None
@@ -61,6 +71,108 @@ def configure_quiet_logging() -> None:
     os.environ.setdefault("HF_HUB_DISABLE_PROGRESS_BARS", "1")
 
 
+def _parse_nonnegative_int(value: Optional[str]) -> Optional[int]:
+    if value is None:
+        return None
+    try:
+        parsed = int(value)
+    except (TypeError, ValueError):
+        return None
+    return parsed if parsed >= 0 else None
+
+
+def _first_mpi_env_pair() -> tuple[Optional[int], Optional[int]]:
+    for rank_name, size_name in _MPI_ENV_PAIRS:
+        rank = _parse_nonnegative_int(os.environ.get(rank_name))
+        world_size = _parse_nonnegative_int(os.environ.get(size_name))
+        if rank is not None and world_size is not None and world_size > 1 and rank < world_size:
+            return rank, world_size
+    return None, None
+
+
+def _json_rank_count_from_env(name: str) -> Optional[int]:
+    value = os.environ.get(name)
+    if not value:
+        return None
+    try:
+        if value.lstrip().startswith(("[", "{")):
+            data = json.loads(value)
+        else:
+            with open(value, "r") as f:
+                data = json.load(f)
+    except (OSError, json.JSONDecodeError):
+        return None
+    if isinstance(data, list):
+        return len(data)
+    if isinstance(data, dict) and isinstance(data.get("hosts"), list):
+        return len(data["hosts"])
+    return None
+
+
+def mlx_distributed_info() -> tuple[bool, int, Optional[int]]:
+    """Return launch-context metadata without initializing MLX distributed."""
+    rank = _parse_nonnegative_int(os.environ.get("MLX_RANK"))
+    world_size = _parse_nonnegative_int(os.environ.get("MLX_WORLD_SIZE"))
+    if rank is not None:
+        if (
+            world_size is not None
+            and world_size > 1
+            and rank < world_size
+            and os.environ.get("NCCL_HOST_IP")
+            and os.environ.get("NCCL_PORT")
+        ):
+            return True, rank, world_size
+        inferred_size = _json_rank_count_from_env("MLX_HOSTFILE")
+        if inferred_size is not None and inferred_size > 1 and rank < inferred_size:
+            return True, rank, inferred_size
+        inferred_size = _json_rank_count_from_env("MLX_IBV_DEVICES")
+        if (
+            inferred_size is not None
+            and inferred_size > 1
+            and rank < inferred_size
+            and os.environ.get("MLX_JACCL_COORDINATOR")
+        ):
+            return True, rank, inferred_size
+        return False, 0, None
+
+    mpi_rank, mpi_world_size = _first_mpi_env_pair()
+    return mpi_rank is not None, mpi_rank or 0, mpi_world_size
+
+
+def mlx_distributed_uses_mpi() -> bool:
+    """Whether the current distributed context was launched through MPI."""
+    return (
+        _parse_nonnegative_int(os.environ.get("MLX_RANK")) is None
+        and _first_mpi_env_pair()[0] is not None
+    )
+
+
+@contextmanager
+def quiet_if_nonzero_mlx_rank():
+    """Silence parent and child-process stdout/stderr on nonzero ranks."""
+    if mlx_distributed_info()[1] == 0:
+        yield
+        return
+
+    sys.stdout.flush()
+    sys.stderr.flush()
+    saved_stdout_fd = os.dup(1)
+    saved_stderr_fd = os.dup(2)
+    with open(os.devnull, "w") as devnull:
+        try:
+            os.dup2(devnull.fileno(), 1)
+            os.dup2(devnull.fileno(), 2)
+            with redirect_stdout(devnull), redirect_stderr(devnull):
+                yield
+        finally:
+            sys.stdout.flush()
+            sys.stderr.flush()
+            os.dup2(saved_stdout_fd, 1)
+            os.dup2(saved_stderr_fd, 2)
+            os.close(saved_stdout_fd)
+            os.close(saved_stderr_fd)
+
+
 def visible_text(text: str, show_thinking: bool) -> str:
     if show_thinking:
         return text
@@ -120,6 +232,21 @@ def collect_stream(stream, show_thinking: bool) -> str:
     return visible_text(raw, show_thinking)
 
 
+def raise_on_streamed_error(stream):
+    # Match real backend errors by type (GenStreamError), not the "Error:" text
+    # prefix, so a completion whose text opens with "Error:" is not misread as a
+    # failure that aborts a distributed run.
+    try:
+        ensure_studio_backend_path()
+        from core.inference.orchestrator import GenStreamError
+    except Exception:
+        GenStreamError = None
+    for chunk in stream:
+        if GenStreamError is not None and isinstance(chunk, GenStreamError):
+            raise RuntimeError(str(chunk)[len(_STREAMED_ERROR_PREFIX) :].strip() or "Unknown error")
+        yield chunk
+
+
 def render_columns(
     left_label: str,
     left_text: str,
@@ -200,6 +327,19 @@ class ChatBackend:
         except Exception:
             pass
 
+    def share_distributed_object(
+        self,
+        obj,
+        *,
+        timeout = 300.0,
+    ):
+        if self._kind != "unsloth" or not hasattr(self._backend, "share_distributed_object"):
+            raise RuntimeError(
+                "Distributed MLX chat requires the Unsloth MLX backend; "
+                f"backend '{self._kind}' cannot broadcast chat turns."
+            )
+        return self._backend.share_distributed_object(obj, timeout = timeout)
+
 
 def resolve_model_config(model: str, *, hf_token: Optional[str]):
     ensure_studio_backend_path()
@@ -293,36 +433,59 @@ def load_chat_backend(
     fresh_backend uses a private orchestrator so a second model (compare's
     base column) can run alongside the main one.
     """
-    if model_config is None:
-        model_config = resolve_model_config(model, hf_token = hf_token)
+    with quiet_if_nonzero_mlx_rank():
+        is_mlx_distributed, rank, _world_size = mlx_distributed_info()
+        if model_config is None:
+            model_config = resolve_model_config(model, hf_token = hf_token)
 
-    typer.echo(f"Loading {model}", err = True)
+        if is_mlx_distributed and model_config.is_gguf:
+            if rank == 0:
+                typer.echo(
+                    "Distributed MLX inference does not support GGUF/llama.cpp models. "
+                    "Use a non-GGUF MLX model under mlx.launch, or run GGUF without "
+                    "mlx.launch.",
+                    err = True,
+                )
+            raise typer.Exit(code = 1)
 
-    if model_config.is_gguf:
-        return _load_gguf_backend(
-            model_config,
-            hf_token = hf_token,
-            max_seq_length = max_seq_length,
-            tensor_parallel = tensor_parallel,
-            llama_extra_args = llama_extra_args,
-        )
+        if rank == 0:
+            typer.echo(f"Loading {model}", err = True)
 
-    if fresh_backend:
-        ensure_studio_backend_path()
-        from core.inference import InferenceOrchestrator
-        backend = InferenceOrchestrator()
-    else:
-        ensure_studio_backend_path()
-        from core.inference import get_inference_backend
-        backend = get_inference_backend()
-    if not backend.load_model(
-        config = model_config,
-        max_seq_length = max_seq_length,
-        load_in_4bit = load_in_4bit,
-        hf_token = hf_token,
-    ):
-        typer.echo("Model load failed", err = True)
-        raise typer.Exit(code = 1)
+        if model_config.is_gguf:
+            return _load_gguf_backend(
+                model_config,
+                hf_token = hf_token,
+                max_seq_length = max_seq_length,
+                tensor_parallel = tensor_parallel,
+                llama_extra_args = llama_extra_args,
+            )
+
+        if fresh_backend:
+            ensure_studio_backend_path()
+            from core.inference import InferenceOrchestrator
+            backend = InferenceOrchestrator()
+        else:
+            ensure_studio_backend_path()
+            from core.inference import get_inference_backend
+            backend = get_inference_backend()
+        try:
+            loaded = backend.load_model(
+                config = model_config,
+                max_seq_length = max_seq_length,
+                load_in_4bit = load_in_4bit,
+                hf_token = hf_token,
+                tensor_parallel = tensor_parallel,
+                mlx_distributed = is_mlx_distributed,
+            )
+        except Exception as exc:
+            if not is_mlx_distributed:
+                raise
+            if rank == 0:
+                typer.echo(str(exc) or "Model load failed", err = True)
+            raise typer.Exit(code = 1)
+        if not loaded:
+            typer.echo("Model load failed", err = True)
+            raise typer.Exit(code = 1)
     return ChatBackend("unsloth", backend)
 
 
diff --git a/unsloth_cli/commands/chat.py b/unsloth_cli/commands/chat.py
index d3bfbbf96b..bc4a72f36c 100644
--- a/unsloth_cli/commands/chat.py
+++ b/unsloth_cli/commands/chat.py
@@ -1,6 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
+import sys
 from typing import List, Optional
 
 import typer
@@ -12,6 +13,10 @@ from unsloth_cli._inference import (
     connect_studio_server,
     ensure_studio_backend_path,
     load_chat_backend,
+    mlx_distributed_info,
+    mlx_distributed_uses_mpi,
+    quiet_if_nonzero_mlx_rank,
+    raise_on_streamed_error,
     render_columns,
     resolve_model_config,
     stream_markdown,
@@ -107,6 +112,20 @@ def _compare_needs_second_model() -> bool:
         return False
 
 
+def _drain_available_stdin() -> None:
+    """Drain already-buffered launcher stdin on nonzero distributed ranks."""
+    try:
+        import os
+        from select import select
+
+        fd = sys.stdin.fileno()
+        while select([fd], [], [], 0)[0]:
+            if not os.read(fd, 8192):
+                break
+    except Exception:
+        return
+
+
 def _pick_trained_model(console) -> str:
     ensure_studio_backend_path()
     from utils.models import scan_trained_models
@@ -158,7 +177,8 @@ def chat(
         "--tensor-parallel/--no-tensor-parallel",
         help = (
             "Split a GGUF across GPUs by tensor (--split-mode tensor) instead "
-            "of by layer. Ignored for non-GGUF models."
+            "of by layer. Under non-MPI mlx.launch, select MLX tensor "
+            "parallel mode instead of pipeline mode."
         ),
     ),
     llama_extra_args: Optional[List[str]] = typer.Option(
@@ -195,15 +215,43 @@ def chat(
 
     console = Console()
     err = Console(stderr = True)
+    is_mlx_distributed, rank, _world_size = mlx_distributed_info()
+    should_print = rank == 0
+
+    if is_mlx_distributed and mlx_distributed_uses_mpi():
+        if should_print:
+            err.print(
+                "Distributed `unsloth chat` with MPI needs rank-0 prompt broadcast, "
+                "which is not enabled yet. Use a non-MPI MLX launcher backend "
+                "such as ring/JACCL for now.",
+                style = "red",
+                markup = False,
+            )
+        raise typer.Exit(code = 1)
 
     if model is None:
+        if is_mlx_distributed:
+            if should_print:
+                err.print(
+                    "Distributed `unsloth chat` requires an explicit model id or path.",
+                    style = "red",
+                    markup = False,
+                )
+            raise typer.Exit(code = 1)
         model = _pick_trained_model(console)
 
     # Resolve first so --compare can be rejected before the slow load.
-    model_config = resolve_model_config(model, hf_token = hf_token)
+    with quiet_if_nonzero_mlx_rank():
+        model_config = resolve_model_config(model, hf_token = hf_token)
     compare_blocked = _compare_blocked_reason(model_config)
+    if is_mlx_distributed:
+        compare_blocked = (
+            "distributed MLX chat does not support compare mode yet because it "
+            "would need a second distributed worker group on the same ranks"
+        )
     if compare and compare_blocked:
-        err.print(f"--compare unavailable: {compare_blocked}", style = "red", markup = False)
+        if should_print:
+            err.print(f"--compare unavailable: {compare_blocked}", style = "red", markup = False)
         raise typer.Exit(code = 1)
 
     load_opts = dict(
@@ -215,9 +263,11 @@ def chat(
     )
 
     # Prefer a running Studio server: instant starts, model shared with the UI.
-    chat_backend = None if no_server else connect_studio_server(model, **load_opts)
+    chat_backend = (
+        None if (no_server or is_mlx_distributed) else connect_studio_server(model, **load_opts)
+    )
     server_mode = chat_backend is not None
-    if server_mode:
+    if server_mode and should_print:
         console.print(
             "(Studio server connected — model stays warm after /exit)",
             style = "bright_black",
@@ -242,23 +292,26 @@ def chat(
             return True
         base_id = model_config.base_model
         if not base_id:
-            console.print(
-                "(compare unavailable: this adapter doesn't record its base model)",
-                style = "yellow",
-            )
+            if should_print:
+                console.print(
+                    "(compare unavailable: this adapter doesn't record its base model)",
+                    style = "yellow",
+                )
             return False
-        console.print(
-            f"(loading base model {base_id} for compare — keeps two models in memory)",
-            style = "bright_black",
-            markup = False,
-        )
+        if should_print:
+            console.print(
+                f"(loading base model {base_id} for compare — keeps two models in memory)",
+                style = "bright_black",
+                markup = False,
+            )
         try:
             # Use the same precision as the tuned model for fair comparison
             base_load_opts = dict(load_opts)  # Copy original options
             base_load_opts["load_in_4bit"] = _get_base_load_in_4bit(model_config)
             base_backend = load_chat_backend(base_id, fresh_backend = True, **base_load_opts)
         except Exception as exc:
-            err.print(f"(base model load failed: {exc})", style = "red", markup = False)
+            if should_print:
+                err.print(f"(base model load failed: {exc})", style = "red", markup = False)
             return False
         return True
 
@@ -267,7 +320,7 @@ def chat(
 
     def generate(backend = None, use_adapter = None):
         # Reads messages and show_thinking live, so /reset and /think apply.
-        return (backend or chat_backend).stream(
+        stream = (backend or chat_backend).stream(
             messages,
             system_prompt = system_prompt,
             temperature = temperature,
@@ -278,22 +331,48 @@ def chat(
             enable_thinking = show_thinking,
             use_adapter = use_adapter,
         )
+        return raise_on_streamed_error(stream) if is_mlx_distributed else stream
 
-    console.print()
-    console.print(f"Chatting with {name}", style = "bold green", markup = False)
-    console.print(_HELP, style = "bright_black")
+    if should_print:
+        console.print()
+        console.print(f"Chatting with {name}", style = "bold green", markup = False)
+        console.print(_HELP, style = "bright_black")
 
     # legacy_windows: pre-VT consoles print raw ANSI as ←[1;36m garbage.
-    you_prompt = _you_prompt(console.is_terminal and not console.legacy_windows)
+    you_prompt = (
+        _you_prompt(console.is_terminal and not console.legacy_windows) if should_print else ""
+    )
     assistant_label = "[bold magenta]Assistant:[/bold magenta]"
 
     try:
         while True:
-            try:
-                user = input(you_prompt).strip()
-            except (EOFError, KeyboardInterrupt):
-                console.print()
-                break
+            if should_print:
+                try:
+                    user = input(you_prompt).strip()
+                except (EOFError, KeyboardInterrupt):
+                    if should_print:
+                        console.print()
+                    user = "/exit"
+                turn = {"type": "turn", "text": user}
+            else:
+                turn = None
+
+            if is_mlx_distributed:
+                try:
+                    turn = chat_backend.share_distributed_object(turn, timeout = None)
+                    if not should_print:
+                        _drain_available_stdin()
+                except Exception as exc:
+                    if should_print:
+                        err.print(
+                            f"\n(error sharing chat turn: {exc})",
+                            style = "red",
+                            markup = False,
+                        )
+                    raise typer.Exit(code = 1)
+            if not turn:
+                continue
+            user = str(turn.get("text", "")).strip()
 
             if not user:
                 continue
@@ -301,55 +380,69 @@ def chat(
                 break
             if user == "/reset":
                 messages = []
-                console.print("(history cleared)", style = "bright_black")
+                if should_print:
+                    console.print("(history cleared)", style = "bright_black")
                 continue
             if user == "/think":
                 show_thinking = not show_thinking
-                state = "on" if show_thinking else "off"
-                console.print(f"(thinking {state})", style = "bright_black")
+                if should_print:
+                    state = "on" if show_thinking else "off"
+                    console.print(f"(thinking {state})", style = "bright_black")
                 continue
             if user == "/compare":
                 if compare_blocked:
-                    console.print(f"(compare unavailable: {compare_blocked})", style = "yellow")
+                    if should_print:
+                        console.print(f"(compare unavailable: {compare_blocked})", style = "yellow")
                     continue
                 if not compare_mode and dual_compare and not load_base_for_compare():
                     continue
                 compare_mode = not compare_mode
-                state = "on" if compare_mode else "off"
-                console.print(f"(compare {state})", style = "bright_black")
+                if should_print:
+                    state = "on" if compare_mode else "off"
+                    console.print(f"(compare {state})", style = "bright_black")
                 continue
             if user in ("/help", "/?"):
-                console.print(_HELP, style = "bright_black")
+                if should_print:
+                    console.print(_HELP, style = "bright_black")
                 continue
 
             messages.append({"role": "user", "content": user})
 
             try:
                 if compare_mode:
-                    console.print("(comparing base vs tuned…)", style = "bright_black")
+                    if should_print:
+                        console.print("(comparing base vs tuned…)", style = "bright_black")
                     if dual_compare:
                         base_text = collect_stream(generate(backend = base_backend), show_thinking)
                         tuned_text = collect_stream(generate(), show_thinking)
                     else:
                         base_text = collect_stream(generate(use_adapter = False), show_thinking)
                         tuned_text = collect_stream(generate(use_adapter = True), show_thinking)
-                    console.print()
-                    render_columns(
-                        "base", base_text, f"{name} (tuned)", tuned_text, console = console
-                    )
+                    if should_print:
+                        console.print()
+                        render_columns(
+                            "base", base_text, f"{name} (tuned)", tuned_text, console = console
+                        )
                     # History continues as the tuned model; base is just the reference.
                     answer = tuned_text
                 else:
-                    console.print(assistant_label)
-                    answer = stream_markdown(generate(), show_thinking, console = console)
+                    if should_print:
+                        console.print(assistant_label)
+                        answer = stream_markdown(generate(), show_thinking, console = console)
+                    else:
+                        answer = collect_stream(generate(), show_thinking)
             except KeyboardInterrupt:
                 # Ctrl-C aborts this answer only; drop the unanswered turn.
-                console.print("\n(interrupted)", style = "bright_black")
+                if should_print:
+                    console.print("\n(interrupted)", style = "bright_black")
                 messages.pop()
                 continue
             except Exception as exc:
-                err.print(f"\n(error: {exc})", style = "red", markup = False)
+                if should_print:
+                    err.print(f"\n(error: {exc})", style = "red", markup = False)
                 messages.pop()
+                if is_mlx_distributed:
+                    raise typer.Exit(code = 1)
                 continue
 
             messages.append(
@@ -359,4 +452,5 @@ def chat(
         chat_backend.close()
         if base_backend is not None:
             base_backend.close()
-        err.print("\nBye.", style = "bright_black")
+        if should_print:
+            err.print("\nBye.", style = "bright_black")
diff --git a/unsloth_cli/commands/inference.py b/unsloth_cli/commands/inference.py
index 1401fa9e9e..84a126163e 100644
--- a/unsloth_cli/commands/inference.py
+++ b/unsloth_cli/commands/inference.py
@@ -6,9 +6,13 @@ from typing import List, Optional
 import typer
 
 from unsloth_cli._inference import (
+    collect_stream,
     configure_quiet_logging,
     connect_studio_server,
     load_chat_backend,
+    mlx_distributed_info,
+    mlx_distributed_uses_mpi,
+    raise_on_streamed_error,
     stream_to_stdout,
 )
 
@@ -36,7 +40,8 @@ def inference(
         "--tensor-parallel/--no-tensor-parallel",
         help = (
             "Split a GGUF across GPUs by tensor (--split-mode tensor) instead "
-            "of by layer. Ignored for non-GGUF models."
+            "of by layer. Under non-MPI mlx.launch, select MLX tensor "
+            "parallel mode instead of pipeline mode."
         ),
     ),
     llama_extra_args: Optional[List[str]] = typer.Option(
@@ -69,8 +74,20 @@ def inference(
     if not verbose:
         configure_quiet_logging()
 
-    # A running Studio server keeps the model warm between runs, which is
-    # exactly what a one-shot command wants.
+    is_mlx_distributed, rank, _world_size = mlx_distributed_info()
+    if is_mlx_distributed and mlx_distributed_uses_mpi():
+        if rank == 0:
+            typer.echo(
+                "Distributed `unsloth inference` with MPI is not supported by "
+                "the current subprocess backend. Use a non-MPI MLX launcher "
+                "backend such as ring/JACCL for now.",
+                err = True,
+            )
+        raise typer.Exit(code = 1)
+
+    # A running Studio server keeps the model warm between runs. Under
+    # mlx.launch, every rank must enter the local MLX path instead of rank 0
+    # alone talking to a server.
     load_opts = dict(
         hf_token = hf_token,
         max_seq_length = max_seq_length,
@@ -78,7 +95,9 @@ def inference(
         tensor_parallel = tensor_parallel,
         llama_extra_args = llama_extra_args,
     )
-    chat_backend = None if no_server else connect_studio_server(model, **load_opts)
+    chat_backend = (
+        None if (no_server or is_mlx_distributed) else connect_studio_server(model, **load_opts)
+    )
     if chat_backend is None:
         chat_backend = load_chat_backend(model, **load_opts)
     try:
@@ -92,7 +111,23 @@ def inference(
             repetition_penalty = repetition_penalty,
             enable_thinking = think,
         )
-        typer.echo("Assistant:")
-        stream_to_stdout(stream, show_thinking = think)
+        if is_mlx_distributed:
+            stream = raise_on_streamed_error(stream)
+        if rank == 0:
+            typer.echo("Assistant:")
+            try:
+                stream_to_stdout(stream, show_thinking = think)
+            except RuntimeError as exc:
+                if not is_mlx_distributed:
+                    raise
+                typer.echo(f"Error: {exc}", err = True)
+                raise typer.Exit(code = 1)
+        else:
+            try:
+                collect_stream(stream, show_thinking = think)
+            except RuntimeError:
+                if not is_mlx_distributed:
+                    raise
+                raise typer.Exit(code = 1)
     finally:
         chat_backend.close()
diff --git a/unsloth_cli/commands/start.py b/unsloth_cli/commands/start.py
index 1895125b11..48c0aca34b 100644
--- a/unsloth_cli/commands/start.py
+++ b/unsloth_cli/commands/start.py
@@ -44,12 +44,31 @@ _CODEX_PROFILE = "unsloth_api"
 _CODEX_ENV_KEY = "UNSLOTH_STUDIO_AUTH_TOKEN"
 _HERMES_ENV_KEY = "UNSLOTH_API_KEY"
 _HERMES_PROVIDER = "unsloth"
+# Skip the installer's interactive setup wizard: `unsloth start hermes` runs
+# this hint unattended and then writes its own session-scoped Hermes config, so
+# the wizard's global API-key/model prompts would block the launch and point the
+# user at a different (global) provider than the one Unsloth just configured.
+# Both installers expose a skip flag: `-SkipSetup` (PowerShell) and
+# `--skip-setup` (POSIX; passed to the piped script via `bash -s --`).
+_HERMES_WINDOWS_INSTALL_HINT = (
+    "& ([scriptblock]::Create((irm https://hermes-agent.nousresearch.com/install.ps1))) -SkipSetup"
+)
+_HERMES_POSIX_INSTALL_HINT = (
+    "curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent"
+    "/main/scripts/install.sh | bash -s -- --skip-setup"
+)
 # Hermes refuses to initialize when the model window is under 64,000 tokens; its
 # error message points at the model.context_length / auxiliary.compression
 # overrides in config.yaml. write_hermes_config claims this value for smaller
 # windows and scales the compaction threshold back down to the real window.
 _HERMES_MIN_CONTEXT = 65536
 _PI_PROVIDER = "unsloth"
+# OpenCode selects a model by "<providerID>/<modelID>" and honors a user
+# disabled_providers list. Register the session provider under a dedicated id a
+# user's disable list would never target, so the model is always selectable
+# without the wrapper having to reconstruct (and override) OpenCode's full,
+# multi-layer disabled_providers resolution.
+_OPENCODE_PROVIDER = "unsloth-studio"
 _PROVIDER_HEADER = f"[model_providers.{_CODEX_PROFILE}]"
 _PASSTHROUGH = {"allow_extra_args": True, "ignore_unknown_options": True}
 _CLAUDE_ENV_UNSET = ("ANTHROPIC_API_KEY", "CLAUDE_CODE_OAUTH_TOKEN")
@@ -114,6 +133,21 @@ _YOLO_OPTION = typer.Option(
         "flag/config. Any of the three spellings works for any agent."
     ),
 )
+_PERSIST_OPTION = typer.Option(
+    False,
+    "--persist/--no-persist",
+    help = (
+        "Keep this agent's Unsloth-managed session dir so you can resume it later. "
+        "codex/openclaw/hermes/pi have their whole home relocated into an Unsloth dir "
+        "that is a throwaway temp dir (wiped on exit) by default; with --persist it "
+        "lives under the Unsloth agents dir and survives, so their own resume can reopen "
+        "it. claude and opencode keep sessions in your own stores (~/.claude, "
+        "~/.local/share/opencode), so they already resume regardless. To reopen a "
+        "session, pass the agent's own resume command through, e.g. "
+        "`unsloth start codex --persist resume` or `claude --resume <id>`; those flow to "
+        "the agent unchanged."
+    ),
+)
 
 # Per-agent CLI flag for "run tools without prompting". opencode and openclaw have no
 # such flag (config only) and are handled in their config writers, so they are absent.
@@ -132,6 +166,10 @@ def _yolo_command_flags(agent: str, yolo: bool) -> list:
     return _YOLO_COMMAND_FLAGS.get(agent, []) if yolo else []
 
 
+def _hermes_install_hint() -> str:
+    return _HERMES_WINDOWS_INSTALL_HINT if os.name == "nt" else _HERMES_POSIX_INSTALL_HINT
+
+
 class LoadOptions(NamedTuple):
     """Model-load knobs forwarded to /api/inference/load when --model triggers a load."""
 
@@ -574,6 +612,59 @@ def _loaded_models(base: str, key: str) -> list:
     return _http_json("GET", f"{base}/v1/models", key, error = "Couldn't list models").get("data", [])
 
 
+_HF_REPO_ID_SEGMENT_RE = re.compile(r"^[A-Za-z0-9._-]+$")
+
+
+def _is_hub_model_id(value: object) -> bool:
+    if not isinstance(value, str):
+        return False
+    text = value.strip()
+    if "\\" in text:
+        return False
+    if text.startswith(("/", "./", "../", "~")):
+        return False
+    if len(text) >= 2 and text[1] == ":" and text[0].isalpha():
+        return False
+    # A hub id is exactly "namespace/name" over a restricted charset. Anything with
+    # extra path segments (e.g. a server-side relative path such as
+    # models/Llama/Foo.gguf on a remote Studio) is not a hub id and must not be
+    # casefold-matched against a differently cased path on a case-sensitive
+    # filesystem. This is host independent, unlike the existence probe below which
+    # cannot see a path that only exists on the server.
+    parts = text.split("/")
+    if len(parts) != 2:
+        return False
+    if any(part in ("", ".", "..") or not _HF_REPO_ID_SEGMENT_RE.match(part) for part in parts):
+        return False
+    try:
+        if Path(os.path.expanduser(text)).exists():
+            return False
+    except OSError:
+        return False
+    return True
+
+
+def _model_id_matches(
+    actual: object,
+    requested: object,
+    *,
+    allow_casefold: bool = True,
+) -> bool:
+    if actual == requested:
+        return True
+    # Case-insensitive matching is only safe when the local existence probe in
+    # _is_hub_model_id is authoritative, i.e. against a loopback Studio on this host.
+    # Against a remote Studio a two-segment string is indistinguishable from a
+    # server-side relative path (e.g. Models/Foo vs models/foo), so casefolding it
+    # could attach to the wrong model on a case-sensitive server; defer to an exact
+    # match there and let the load endpoint resolve the requested path.
+    if not allow_casefold:
+        return False
+    if not (_is_hub_model_id(actual) and _is_hub_model_id(requested)):
+        return False
+    return str(actual).casefold() == str(requested).casefold()
+
+
 def _resolve_model(
     base: str,
     key: str,
@@ -581,6 +672,9 @@ def _resolve_model(
     load: LoadOptions = LoadOptions(),
 ) -> dict:
     models = _loaded_models(base, key)
+    # Only casefold-match ids against a loopback Studio, where _is_hub_model_id's
+    # local existence probe can actually reject a server-side path; see the note there.
+    allow_casefold = is_loopback_url(base)
     # /v1/models reports the model id but not the active GGUF variant or runtime load
     # settings, so an id match alone can hide the wrong quant (Q8_0 serving while the
     # user asked for UD-Q4_K_XL). When the user passed any explicit load knob, defer to
@@ -590,10 +684,21 @@ def _resolve_model(
     load_has_overrides = bool(
         load.gguf_variant or load.max_seq_length or not load.load_in_4bit or load.tensor_parallel
     )
+    # /v1/models also lists cached-but-unloaded catalog entries (loaded == False);
+    # matching one would skip /api/inference/load and leave the agent pointed at a
+    # model that is not resident, so only attach to an entry that is actually loaded.
     match = (
         None
         if requested and load_has_overrides
-        else next((m for m in models if m["id"] == requested), None)
+        else next(
+            (
+                m
+                for m in models
+                if _model_id_matches(m.get("id"), requested, allow_casefold = allow_casefold)
+                and m.get("loaded") is not False
+            ),
+            None,
+        )
     )
     if requested and match is None:
         typer.echo(
@@ -628,7 +733,16 @@ def _resolve_model(
         if isinstance(loaded, dict):
             wanted |= {loaded.get("model"), loaded.get("display_name")} - {None}
         models = _loaded_models(base, key)
-        match = next((m for m in models if m["id"] in wanted), None)
+        match = next(
+            (
+                m
+                for m in models
+                if any(
+                    _model_id_matches(m.get("id"), w, allow_casefold = allow_casefold) for w in wanted
+                )
+            ),
+            None,
+        )
     if match is not None:
         return match
     if requested:
@@ -842,6 +956,60 @@ def _print_env(
     typer.echo(" ".join((*inline, shlex.join(command))))
 
 
+def _refresh_windows_path() -> None:
+    # Merge Windows registry PATH hives after the current process PATH so a
+    # freshly installed agent is visible without changing existing precedence.
+    if os.name != "nt":
+        return
+    try:
+        import winreg
+    except Exception:
+        return
+
+    entries = []
+    seen = set()
+
+    def add_path(value: str) -> bool:
+        added = False
+        for entry in str(value).split(os.pathsep):
+            entry = entry.strip()
+            if not entry:
+                continue
+            key = os.path.normcase(entry).casefold()
+            if key in seen:
+                continue
+            seen.add(key)
+            entries.append(entry)
+            added = True
+        return added
+
+    add_path(os.environ.get("PATH", ""))
+    added_registry = False
+    hives = (
+        (winreg.HKEY_CURRENT_USER, "Environment"),
+        (
+            winreg.HKEY_LOCAL_MACHINE,
+            r"SYSTEM\CurrentControlSet\Control\Session Manager\Environment",
+        ),
+    )
+    for root, sub in hives:
+        try:
+            with winreg.OpenKey(root, sub) as key:
+                value, _ = winreg.QueryValueEx(key, "Path")
+        except OSError:
+            continue
+        if value:
+            added_registry = add_path(os.path.expandvars(str(value))) or added_registry
+    if added_registry:
+        os.environ["PATH"] = os.pathsep.join(entries)
+
+
+def _install_source(install_hint: str) -> Optional[str]:
+    """The first http(s) URL an install hint fetches, or None (e.g. an npm install)."""
+    match = re.search(r"https?://[^\s'\")]+", install_hint)
+    return match.group(0) if match else None
+
+
 def _install_agent(name: str, install_hint: str) -> Optional[str]:
     # Missing agent under --launch: offer to run its documented install command, then
     # re-resolve it on PATH. Consent-based (we never auto-run a remote install script
@@ -850,7 +1018,18 @@ def _install_agent(name: str, install_hint: str) -> Optional[str]:
     if not sys.stdin.isatty():
         return None
     typer.echo(f"`{name}` is not installed.")
-    if not typer.confirm(f"Install it now with `{install_hint}`?", default = False):
+    # Make the supply-chain risk explicit before the prompt: these are the vendors'
+    # own installers (curl | bash, irm | iex, npm), run with the user's privileges,
+    # and nothing checks a signature or hash on the fetched content. Naming the source
+    # turns a blind "yes" into informed consent.
+    source = _install_source(install_hint)
+    warning = (
+        f"This will download and RUN a script from {source} with your privileges"
+        if source
+        else f"This will RUN `{install_hint}` with your privileges"
+    )
+    typer.secho(f"{warning}; there is no signature or hash check.", fg = "yellow", err = True)
+    if not typer.confirm(f"Install `{name}` now with `{install_hint}`?", default = False):
         return None
     # Run each hint through the shell it is written for: PowerShell (irm | iex, or npm)
     # on Windows, /bin/sh (curl | bash, or npm) everywhere else.
@@ -860,6 +1039,9 @@ def _install_agent(name: str, install_hint: str) -> Optional[str]:
         install_command = ["/bin/sh", "-c", install_hint]
     if subprocess.run(install_command).returncode != 0:
         _fail(f"Install command failed. Run it yourself, then re-run: {install_hint}")
+    # The installer just wrote PATH to the registry (Windows); pull it into this
+    # process so the freshly installed agent resolves without a shell restart.
+    _refresh_windows_path()
     executable = shutil.which(name)
     if executable is None:
         _fail(
@@ -966,15 +1148,20 @@ def _agents_config_root() -> Path:
 
 
 @contextlib.contextmanager
-def _session_config(agent: str, launch: bool):
+def _session_config(
+    agent: str,
+    launch: bool,
+    persist: bool = False,
+):
     """Yield a private directory for an agent's session config (never the user's own).
 
-    launch: an ephemeral temp dir removed after the agent process exits, so nothing
-    persists. no-launch: a stable Unsloth-owned dir (the printed recipe is run later
-    on this machine), reused across runs. Either way the user's real ~/.<agent>
-    config is left untouched.
+    launch (default): an ephemeral temp dir removed after the agent process exits, so
+    nothing persists. no-launch: a stable Unsloth-owned dir (the printed recipe is run
+    later on this machine), reused across runs. persist (from --persist): use that same
+    stable dir even for a launch, so the agent's session survives the exit and can be
+    resumed next time. Either way the user's real ~/.<agent> config is left untouched.
     """
-    if launch:
+    if launch and not persist:
         path = Path(tempfile.mkdtemp(prefix = f"unsloth-{agent}-"))
         try:
             yield path
@@ -1116,13 +1303,17 @@ def write_opencode_config(
     config = _read_json_object(path)
     if config is None:
         typer.echo(
-            f"Warning: couldn't parse {path} — add an 'unsloth' provider there "
-            "yourself, or move the file aside and re-run.",
+            f"Warning: couldn't parse {path} — add an '{_OPENCODE_PROVIDER}' provider "
+            "there yourself, or move the file aside and re-run.",
             err = True,
         )
         return {}
     before = json.dumps(config, sort_keys = True)
     config.setdefault("$schema", "https://opencode.ai/config.json")
+    # The session provider is registered under a dedicated id (_OPENCODE_PROVIDER)
+    # that a user's disabled_providers list would never target, so it is always
+    # selectable without this overlay having to reconstruct or override OpenCode's
+    # disabled_providers resolution.
     model_entry = {"name": model["id"]}
     window = model.get("context_length") or model.get("max_context_length")
     if window:
@@ -1131,14 +1322,14 @@ def write_opencode_config(
         # disables OpenCode's auto-compaction; declare the real window (and a sane
         # output cap) so it compacts instead of overflowing the server.
         model_entry["limit"] = {"context": window, "output": min(window // 4, 8192)}
-    _subdict(config, "provider")["unsloth"] = {
+    _subdict(config, "provider")[_OPENCODE_PROVIDER] = {
         "npm": "@ai-sdk/openai-compatible",
         "name": "Unsloth Studio",
         "options": {"baseURL": f"{base}/v1", "apiKey": key},
         "models": {model["id"]: model_entry},
     }
     # OpenCode selects a model by "<providerID>/<modelID>".
-    config["model"] = f"unsloth/{model['id']}"
+    config["model"] = f"{_OPENCODE_PROVIDER}/{model['id']}"
     if window:
         # Compact with ~10% headroom (near 90% full). The fixed 20k-token default
         # buffer over-compacts, or never settles, on a small local context.
@@ -1282,6 +1473,7 @@ def claude(
     tensor_parallel: bool = _TENSOR_PARALLEL_OPTION,
     serve: bool = _SERVE_OPTION,
     yolo: bool = _YOLO_OPTION,
+    persist: bool = _PERSIST_OPTION,
 ):
     """Point Claude Code at the running Studio server and start it."""
     base, key, entry = _connect(
@@ -1326,6 +1518,9 @@ def claude(
     # --yolo (or its aliases) maps to Claude's own --dangerously-skip-permissions.
     # IS_SANDBOX is left unset on purpose: Claude refuses bypass mode as root unless a
     # sandbox is detected, and we don't want to falsely claim one on the user's host.
+    # claude keeps its history in ~/.claude/projects, which --settings/env never
+    # relocate, so a session already survives exit; resume it with `claude --continue`
+    # or `--resume <id>` passed through.
     command = [
         "claude",
         "--model",
@@ -1362,6 +1557,7 @@ def codex(
     tensor_parallel: bool = _TENSOR_PARALLEL_OPTION,
     serve: bool = _SERVE_OPTION,
     yolo: bool = _YOLO_OPTION,
+    persist: bool = _PERSIST_OPTION,
 ):
     """Point OpenAI Codex at the running Studio server and start it."""
     base, key, entry = _connect(
@@ -1387,7 +1583,7 @@ def codex(
         *_yolo_command_flags("codex", yolo),
         *ctx.args,
     ]
-    with _session_config("codex", launch) as home:
+    with _session_config("codex", launch, persist = persist) as home:
         write_codex_config(base, entry, home)
         env = {_CODEX_ENV_KEY: key, "CODEX_HOME": str(home)}
         _run(base, entry, env, command, launch = launch, install_hint = "npm install -g @openai/codex")
@@ -1405,6 +1601,7 @@ def openclaw(
     tensor_parallel: bool = _TENSOR_PARALLEL_OPTION,
     serve: bool = _SERVE_OPTION,
     yolo: bool = _YOLO_OPTION,
+    persist: bool = _PERSIST_OPTION,
 ):
     """Point OpenClaw at the running Studio server and start it."""
     base, key, entry = _connect(
@@ -1414,13 +1611,23 @@ def openclaw(
         serve = serve,
         launch = launch,
     )
-    command = ["openclaw", *ctx.args]
+    openclaw_args = list(ctx.args)
+    # Default a bare `unsloth start openclaw` to the local TUI. Anything the caller
+    # passes through is forwarded verbatim so OpenClaw parses it under its own grammar
+    # (openclaw [global-flags] <command> [options]): an explicit subcommand, a global
+    # flag that must precede the command such as --profile/--dev, or a tui option. We
+    # cannot reinterpret those safely because a leading "--flag value" is ambiguous
+    # between a global (`--profile test`) and a tui option (`--message hi`); prepending
+    # `tui --local` would break the global form, so only the empty case is defaulted.
+    if not openclaw_args:
+        openclaw_args = ["tui", "--local"]
+    command = ["openclaw", *openclaw_args]
     install_hint = (
         "iwr -useb https://openclaw.ai/install.ps1 | iex"
         if os.name == "nt"
         else "curl -fsSL https://openclaw.ai/install.sh | bash"
     )
-    with _session_config("openclaw", launch) as cfg:
+    with _session_config("openclaw", launch, persist = persist) as cfg:
         config_path = cfg / "openclaw.json"
         # key lives in the config, not the env; --yolo writes the exec policy here too.
         write_openclaw_config(base, key, entry, config_path, yolo = yolo)
@@ -1441,6 +1648,7 @@ def opencode(
     tensor_parallel: bool = _TENSOR_PARALLEL_OPTION,
     serve: bool = _SERVE_OPTION,
     yolo: bool = _YOLO_OPTION,
+    persist: bool = _PERSIST_OPTION,
 ):
     """Point OpenCode at the running Studio server and start it."""
     base, key, entry = _connect(
@@ -1450,8 +1658,23 @@ def opencode(
         serve = serve,
         launch = launch,
     )
-    command = ["opencode", *ctx.args]
-    with _session_config("opencode", launch) as cfg:
+    opencode_model = f"{_OPENCODE_PROVIDER}/{entry['id']}"
+    # The inline OPENCODE_CONFIG_CONTENT below pins the model in the highest-priority
+    # layer, so the session model is forced without a --model flag. Only add --model for
+    # an interactive bare launch (a convenience so the TUI opens on our model). It is
+    # omitted for passthrough (inserting it before a subcommand can be misparsed) and for
+    # --no-launch, where the printed command is consumed by drivers that append a
+    # subcommand such as `run <prompt>`; a leading --model would land before that
+    # subcommand and break it. Those paths rely on the inline pin instead.
+    if ctx.args:
+        command = ["opencode", *ctx.args]
+    elif launch:
+        command = ["opencode", "--model", opencode_model]
+    else:
+        command = ["opencode"]
+    # opencode keeps sessions in ~/.local/share/opencode (never relocated), so resume
+    # already survives exit; reopen the last one by passing `opencode --continue` through.
+    with _session_config("opencode", launch, persist = persist) as cfg:
         config_path = cfg / "opencode.json"
         # OPENCODE_CONFIG is an overlay (loaded between the user's global and project
         # configs), so this adds the Unsloth provider/model for the session without
@@ -1462,7 +1685,26 @@ def opencode(
         # outranks project config; the API key stays in the private file, never the env.
         # Only --yolo carries a permission here (its allow must win over a project config);
         # a non-yolo session returns no permission, so the project's own rules are honored.
-        inline_config: dict = {"model": f"unsloth/{entry['id']}"}
+        # opencode filters every provider (a config-defined custom one included) through
+        # its enabled_providers allowlist and disabled_providers denylist, and a model pin
+        # does not bypass that gate -- a filtered provider resolves to ModelNotFoundError.
+        # To guarantee the session model loads without reading or modifying the user's real
+        # config, scope THIS session to our provider alone: allowlist _OPENCODE_PROVIDER and
+        # clear the denylist. These arrays are replaced (not merged) by higher layers, so
+        # setting them in the highest-priority inline overlay neutralizes any user allowlist
+        # or denylist for the launch. It is session-only: it lives in OPENCODE_CONFIG_CONTENT
+        # for this invocation and never touches the user's config files, so their normal
+        # `opencode` is unchanged; only this session is limited to the Studio provider.
+        # small_model is opencode's separate model for lightweight tasks; pin it to the
+        # session model too, or a user/project small_model on another (now filtered)
+        # provider would resolve a not-found error mid-session. The session serves one
+        # model, so the session model is the only valid target here anyway.
+        inline_config: dict = {
+            "model": opencode_model,
+            "small_model": opencode_model,
+            "enabled_providers": [_OPENCODE_PROVIDER],
+            "disabled_providers": [],
+        }
         if session_permission:
             inline_config["permission"] = session_permission
         env = {
@@ -1484,6 +1726,7 @@ def hermes(
     tensor_parallel: bool = _TENSOR_PARALLEL_OPTION,
     serve: bool = _SERVE_OPTION,
     yolo: bool = _YOLO_OPTION,
+    persist: bool = _PERSIST_OPTION,
 ):
     """Point Hermes (Nous Research) at the running Studio server and start it."""
     base, key, entry = _connect(
@@ -1494,11 +1737,8 @@ def hermes(
         launch = launch,
     )
     command = ["hermes", *_yolo_command_flags("hermes", yolo), *ctx.args]
-    install_hint = (
-        "curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent"
-        "/main/scripts/install.sh | bash"
-    )
-    with _session_config("hermes", launch) as home:
+    install_hint = _hermes_install_hint()
+    with _session_config("hermes", launch, persist = persist) as home:
         # HERMES_HOME relocates hermes' whole home dir (config.yaml, sessions, state)
         # like CODEX_HOME, so the user's ~/.hermes is left untouched for the session.
         write_hermes_config(base, entry, home / "config.yaml")
@@ -1518,6 +1758,7 @@ def pi(
     tensor_parallel: bool = _TENSOR_PARALLEL_OPTION,
     serve: bool = _SERVE_OPTION,
     yolo: bool = _YOLO_OPTION,
+    persist: bool = _PERSIST_OPTION,
 ):
     """Point Pi (coding agent) at the running Studio server and start it."""
     base, key, entry = _connect(
@@ -1542,7 +1783,7 @@ def pi(
     # --ignore-scripts matches Pi's documented install recipe (its README notes Pi needs
     # no install scripts), so accepting the prompt skips dependency lifecycle scripts.
     install_hint = "npm install -g --ignore-scripts @earendil-works/pi-coding-agent"
-    with _session_config("pi", launch) as home:
+    with _session_config("pi", launch, persist = persist) as home:
         # Pi resolves its config dir from PI_CODING_AGENT_DIR first (getAgentDir() prefers
         # it over $HOME/.pi/agent), so pin it at the session dir: an inherited
         # PI_CODING_AGENT_DIR in the user's shell would otherwise send Pi to their real
diff --git a/unsloth_cli/commands/train.py b/unsloth_cli/commands/train.py
index 9d47a574d5..c52c2344b7 100644
--- a/unsloth_cli/commands/train.py
+++ b/unsloth_cli/commands/train.py
@@ -7,10 +7,42 @@ from typing import Optional
 
 import typer
 
+from unsloth_cli._inference import ensure_studio_backend_path
 from unsloth_cli.config import Config, load_config
 from unsloth_cli.options import add_options_from_config
 
 
+def _should_use_mlx_backend_for_cli() -> bool:
+    ensure_studio_backend_path()
+    from studio.backend.core.training.training import should_use_mlx_training_backend
+    return should_use_mlx_training_backend()
+
+
+def _activate_mlx_transformers(model_name: str, hf_token: Optional[str]) -> None:
+    # Activate before any transformers import: adapter model-type detection imports utils.models.
+    ensure_studio_backend_path()
+    from utils.transformers_version import activate_transformers_for_subprocess
+    try:
+        activate_transformers_for_subprocess(model_name, hf_token)
+    except Exception as exc:
+        typer.echo(f"Warning: failed to activate Transformers sidecar: {exc}", err = True)
+
+
+def _create_cli_trainer(model_name: str, hf_token: Optional[str]):
+    if _should_use_mlx_backend_for_cli():
+        _activate_mlx_transformers(model_name, hf_token)
+        # MLX is torch-free: use the lightweight adapter, not trainer.py (imports torch/unsloth/trl at load).
+        ensure_studio_backend_path()
+        from studio.backend.core.training.training import create_mlx_trainer_adapter
+
+        return create_mlx_trainer_adapter()
+
+    ensure_studio_backend_path()
+    from studio.backend.core.training.trainer import UnslothTrainer
+
+    return UnslothTrainer()
+
+
 @add_options_from_config(Config)
 def train(
     config: Optional[Path] = typer.Option(
@@ -39,6 +71,7 @@ def train(
         typer.echo(f"Error: {e}", err = True)
         raise typer.Exit(code = 2)
 
+    config_overrides = config_overrides or {}
     cfg.apply_overrides(**config_overrides)
 
     # CLI/env tokens take precedence; guard against unresolved typer.Option
@@ -83,9 +116,7 @@ def train(
         )
         raise typer.Exit(code = 2)
 
-    from studio.backend.core.training.trainer import UnslothTrainer
-
-    trainer = UnslothTrainer()
+    trainer = _create_cli_trainer(cfg.model, hf_token)
 
     # Load model (trainer.is_vlm is set after this)
     if not trainer.load_model(
@@ -124,13 +155,20 @@ def train(
 
     try:
         while trainer.training_thread and trainer.training_thread.is_alive():
+            progress = trainer.get_training_progress()
+            if getattr(progress, "error", None):
+                break
             time.sleep(1)
     except KeyboardInterrupt:
         typer.echo("Stopping training (Ctrl+C detected)...")
         trainer.stop_training()
     finally:
         if trainer.training_thread:
-            trainer.training_thread.join()
+            progress = trainer.get_training_progress()
+            if getattr(progress, "error", None):
+                trainer.training_thread.join(timeout = 5)
+            else:
+                trainer.training_thread.join()
 
     final = trainer.get_training_progress()
     if getattr(final, "error", None):
diff --git a/unsloth_cli/tests/test_inference_chat.py b/unsloth_cli/tests/test_inference_chat.py
index 013b7c5a81..56633408fb 100644
--- a/unsloth_cli/tests/test_inference_chat.py
+++ b/unsloth_cli/tests/test_inference_chat.py
@@ -26,6 +26,8 @@ from unsloth_cli._inference import (
     ChatBackend,
     HttpChatBackend,
     collect_stream,
+    mlx_distributed_info,
+    mlx_distributed_uses_mpi,
     render_columns,
     visible_text,
 )
@@ -39,6 +41,16 @@ class _FakeConfig:
     path = None
 
 
+_EXPECTED_MPI_ENV_PAIRS = [
+    ("OMPI_COMM_WORLD_RANK", "OMPI_COMM_WORLD_SIZE"),
+    ("PMI_RANK", "PMI_SIZE"),
+    ("PMIX_RANK", "PMIX_SIZE"),
+    ("MPI_RANK", "MPI_WORLD_SIZE"),
+    ("MV2_COMM_WORLD_RANK", "MV2_COMM_WORLD_SIZE"),
+]
+_IGNORED_DISTRIBUTED_ENV_PAIRS = [("SLURM_PROCID", "SLURM_NTASKS")]
+
+
 def _chat_app():
     cli = typer.Typer()
     cli.command()(chatmod.chat)
@@ -53,6 +65,39 @@ def _inference_app():
     return cli
 
 
+def _clear_mlx_distributed_env(monkeypatch):
+    for name in (
+        "MLX_RANK",
+        "MLX_HOSTFILE",
+        "MLX_WORLD_SIZE",
+        "MLX_IBV_DEVICES",
+        "MLX_JACCL_COORDINATOR",
+        "NCCL_HOST_IP",
+        "NCCL_PORT",
+        *(rank for rank, _size in _EXPECTED_MPI_ENV_PAIRS + _IGNORED_DISTRIBUTED_ENV_PAIRS),
+        *(size for _rank, size in _EXPECTED_MPI_ENV_PAIRS + _IGNORED_DISTRIBUTED_ENV_PAIRS),
+    ):
+        monkeypatch.delenv(name, raising = False)
+
+
+def _set_mlx_nccl_env(
+    monkeypatch,
+    *,
+    rank: str = "0",
+    size: str = "2",
+):
+    monkeypatch.setenv("MLX_RANK", rank)
+    monkeypatch.setenv("MLX_WORLD_SIZE", size)
+    monkeypatch.setenv("NCCL_HOST_IP", "127.0.0.1")
+    monkeypatch.setenv("NCCL_PORT", "12345")
+
+
+@pytest.fixture(autouse = True)
+def _isolate_mlx_distributed_env(monkeypatch):
+    _clear_mlx_distributed_env(monkeypatch)
+    monkeypatch.delenv("HF_TOKEN", raising = False)
+
+
 def test_visible_text_passthrough_when_shown():
     text = "<think>reasoning</think>answer"
     assert visible_text(text, show_thinking = True) == text
@@ -101,6 +146,46 @@ def test_inference_exposes_gguf_runtime_options():
     assert "--llama-extra-arg" in (getattr(extra, "param_decls", None) or [])
 
 
+def test_mlx_distributed_info_reads_launch_env(monkeypatch, tmp_path):
+    _clear_mlx_distributed_env(monkeypatch)
+    assert mlx_distributed_info() == (False, 0, None)
+    assert mlx_distributed_uses_mpi() is False
+
+    monkeypatch.setenv("MLX_RANK", "1")
+    monkeypatch.setenv("MLX_WORLD_SIZE", "2")
+    assert mlx_distributed_info() == (False, 0, None)
+    monkeypatch.setenv("NCCL_HOST_IP", "127.0.0.1")
+    monkeypatch.setenv("NCCL_PORT", "12345")
+    assert mlx_distributed_info() == (True, 1, 2)
+    assert mlx_distributed_uses_mpi() is False
+
+    _clear_mlx_distributed_env(monkeypatch)
+    ring_hostfile = tmp_path / "ring.json"
+    ring_hostfile.write_text('[["127.0.0.1:5000"], ["127.0.0.1:5001"]]\n')
+    monkeypatch.setenv("MLX_RANK", "0")
+    monkeypatch.setenv("MLX_HOSTFILE", str(ring_hostfile))
+    assert mlx_distributed_info() == (True, 0, 2)
+    assert mlx_distributed_uses_mpi() is False
+
+    _clear_mlx_distributed_env(monkeypatch)
+    monkeypatch.setenv("MLX_RANK", "1")
+    monkeypatch.setenv("MLX_IBV_DEVICES", '[["node-a"], ["node-b"]]')
+    monkeypatch.setenv("MLX_JACCL_COORDINATOR", "node-a:12345")
+    assert mlx_distributed_info() == (True, 1, 2)
+    assert mlx_distributed_uses_mpi() is False
+
+    _clear_mlx_distributed_env(monkeypatch)
+    monkeypatch.setenv("OMPI_COMM_WORLD_RANK", "1")
+    monkeypatch.setenv("OMPI_COMM_WORLD_SIZE", "2")
+    assert mlx_distributed_info() == (True, 1, 2)
+    assert mlx_distributed_uses_mpi() is True
+
+    _clear_mlx_distributed_env(monkeypatch)
+    monkeypatch.setenv("MLX_RANK", "bad")
+    monkeypatch.setenv("MLX_WORLD_SIZE", "-3")
+    assert mlx_distributed_info() == (False, 0, None)
+
+
 def test_chat_command_is_registered_with_options():
     params = inspect.signature(chatmod.chat).parameters
     assert "model" in params
@@ -751,7 +836,6 @@ def test_chat_server_mode_compare_loads_base_locally(monkeypatch):
 
     assert result.exit_code == 0, result.output
     assert "(compare on)" in result.output
-    # Only the base model loaded locally, on its own private backend.
     assert base_loads == [("fake/base", True)]
     assert streamed == ["base", "tuned"]
     assert set(closed) == {"http", "base"}
@@ -785,6 +869,248 @@ def test_chat_compare_on_mlx_loads_base_model_side_by_side(monkeypatch):
 
     assert result.exit_code == 0, result.output
     assert loads == [("tuned-run", False), ("fake/base", True)]
-    # Both models answered the turn, via plain generation (no adapter toggle).
     assert ("base", None) in streamed and ("tuned", None) in streamed
     assert set(closed) == {"tuned", "base"}
+
+
+@pytest.mark.parametrize(
+    ("chunk_kind", "expected_exit"),
+    [
+        ("answer", 0),
+        ("model_text_error", 0),
+        ("real_error", 1),
+    ],
+)
+def test_inference_under_mlx_launch_handles_stream(monkeypatch, chunk_kind, expected_exit):
+    from unsloth_cli.commands import inference as infermod
+    from unsloth_cli._inference import ensure_studio_backend_path
+
+    ensure_studio_backend_path()
+    from core.inference.orchestrator import GenStreamError
+
+    if chunk_kind == "answer":
+        chunks = ["answer"]
+    elif chunk_kind == "model_text_error":
+        # Model output whose visible text starts with "Error:" must not abort.
+        chunks = ["Error: printed by the model, not a backend failure"]
+    else:
+        chunks = [GenStreamError("Error: generation failed")]
+
+    loads, closed = [], []
+
+    class _FakeBackend:
+        def stream(self, messages, **kwargs):
+            return iter(chunks)
+
+        def close(self):
+            closed.append(True)
+
+    _set_mlx_nccl_env(monkeypatch, rank = "0")
+    monkeypatch.setattr(
+        infermod,
+        "connect_studio_server",
+        lambda *_a, **_k: (_ for _ in ()).throw(AssertionError("server disabled")),
+    )
+    monkeypatch.setattr(
+        infermod,
+        "load_chat_backend",
+        lambda model, **kwargs: (loads.append((model, kwargs)), _FakeBackend())[1],
+    )
+
+    result = CliRunner().invoke(
+        _inference_app(),
+        ["fake-model", "hello", "--tensor-parallel"],
+    )
+
+    assert result.exit_code == expected_exit, result.output
+    assert loads[0][1]["tensor_parallel"] is True
+    if chunk_kind == "real_error":
+        assert "generation failed" in result.output
+
+
+def test_chat_under_mlx_launch_nonzero_rank_drains_stdin(monkeypatch):
+    drains, closed = [], []
+    turns = iter(
+        [
+            {"type": "turn", "text": "hi"},
+            {"type": "turn", "text": "/exit"},
+        ]
+    )
+
+    class _FakeChatBackend:
+        def share_distributed_object(
+            self,
+            obj,
+            *,
+            timeout = 300.0,
+        ):
+            assert obj is None
+            return next(turns)
+
+        def stream(self, messages, **kwargs):
+            return iter(["hidden"])
+
+        def close(self):
+            closed.append(True)
+
+    _set_mlx_nccl_env(monkeypatch, rank = "1")
+    monkeypatch.setattr(chatmod, "resolve_model_config", lambda *a, **k: _FakeConfig())
+    monkeypatch.setattr(
+        chatmod,
+        "connect_studio_server",
+        lambda *_a, **_k: (_ for _ in ()).throw(AssertionError("server disabled")),
+    )
+    monkeypatch.setattr(chatmod, "load_chat_backend", lambda *a, **k: _FakeChatBackend())
+    monkeypatch.setattr(chatmod, "_compare_needs_second_model", lambda: False)
+    monkeypatch.setattr(chatmod, "_drain_available_stdin", lambda: drains.append(True))
+
+    result = CliRunner().invoke(_chat_app(), ["fake-model"], input = "hi\n/exit\n")
+
+    assert result.exit_code == 0, result.output
+    assert "Chatting with" not in result.output
+    assert drains == [True, True]
+    assert closed == [True]
+
+
+def test_chat_under_mlx_launch_rank0_bypasses_studio_and_prints(monkeypatch):
+    loads, shares, closed = [], [], []
+
+    class _FakeChatBackend:
+        def share_distributed_object(
+            self,
+            obj,
+            *,
+            timeout = 300.0,
+        ):
+            shares.append((obj, timeout))
+            return obj
+
+        def stream(self, messages, **kwargs):
+            return iter(["hello"])
+
+        def close(self):
+            closed.append(True)
+
+    _set_mlx_nccl_env(monkeypatch, rank = "0")
+    monkeypatch.setattr(chatmod, "resolve_model_config", lambda *a, **k: _FakeConfig())
+    monkeypatch.setattr(
+        chatmod,
+        "connect_studio_server",
+        lambda *_a, **_k: (_ for _ in ()).throw(AssertionError("server disabled")),
+    )
+    monkeypatch.setattr(
+        chatmod,
+        "load_chat_backend",
+        lambda model, **kwargs: (loads.append((model, kwargs)), _FakeChatBackend())[1],
+    )
+    monkeypatch.setattr(chatmod, "_compare_needs_second_model", lambda: False)
+
+    result = CliRunner().invoke(
+        _chat_app(),
+        ["fake-model", "--tensor-parallel"],
+        input = "hi\n/exit\n",
+    )
+
+    assert result.exit_code == 0, result.output
+    assert "Chatting with fake-model" in result.output
+    assert "hello" in result.output
+    assert loads and loads[0][0] == "fake-model"
+    assert loads[0][1]["tensor_parallel"] is True
+    assert shares == [
+        ({"type": "turn", "text": "hi"}, None),
+        ({"type": "turn", "text": "/exit"}, None),
+    ]
+
+
+@pytest.mark.parametrize(
+    ("stream_error", "expected_exit"),
+    [("exception", 1), ("chunk", 1), ("model_text", 0)],
+)
+def test_chat_under_mlx_launch_exits_on_generation_error(monkeypatch, stream_error, expected_exit):
+    from unsloth_cli._inference import ensure_studio_backend_path
+
+    ensure_studio_backend_path()
+    from core.inference.orchestrator import GenStreamError
+
+    closed = []
+
+    class _FakeChatBackend:
+        def share_distributed_object(
+            self,
+            obj,
+            *,
+            timeout = 300.0,
+        ):
+            return obj
+
+        def stream(self, messages, **kwargs):
+            if stream_error == "exception":
+                raise RuntimeError("generation failed")
+            if stream_error == "model_text":
+                # Plain model text starting with "Error:" must not abort the run.
+                return iter(["Error: printed by the model"])
+            return iter([GenStreamError("Error: generation failed")])
+
+        def close(self):
+            closed.append(True)
+
+    _set_mlx_nccl_env(monkeypatch, rank = "0")
+    monkeypatch.setattr(chatmod, "resolve_model_config", lambda *a, **k: _FakeConfig())
+    monkeypatch.setattr(
+        chatmod,
+        "connect_studio_server",
+        lambda *_a, **_k: (_ for _ in ()).throw(AssertionError("server disabled")),
+    )
+    monkeypatch.setattr(chatmod, "load_chat_backend", lambda *a, **k: _FakeChatBackend())
+    monkeypatch.setattr(chatmod, "_compare_needs_second_model", lambda: False)
+
+    result = CliRunner().invoke(_chat_app(), ["fake-model"], input = "hi\n/exit\n")
+
+    assert result.exit_code == expected_exit
+    if expected_exit:
+        assert "generation failed" in result.output
+    assert closed == [True]
+
+
+def test_load_chat_backend_forwards_mlx_distributed_options(monkeypatch):
+    import unsloth_cli._inference as inference
+
+    calls = []
+
+    class _FakeBackend:
+        def load_model(self, **kwargs):
+            calls.append(kwargs)
+            return True
+
+    class _FakeModelConfig:
+        is_gguf = False
+
+        @classmethod
+        def from_identifier(cls, **_kwargs):
+            return cls()
+
+    fake_backend = _FakeBackend()
+    fake_inference = types.ModuleType("core.inference")
+    fake_inference.get_inference_backend = lambda: fake_backend
+    fake_utils = types.ModuleType("utils")
+    fake_utils.__path__ = []
+    fake_models = types.ModuleType("utils.models")
+    fake_models.ModelConfig = _FakeModelConfig
+
+    _set_mlx_nccl_env(monkeypatch, rank = "0")
+    monkeypatch.setitem(sys.modules, "core", types.ModuleType("core"))
+    monkeypatch.setitem(sys.modules, "core.inference", fake_inference)
+    monkeypatch.setitem(sys.modules, "utils", fake_utils)
+    monkeypatch.setitem(sys.modules, "utils.models", fake_models)
+    monkeypatch.setattr(inference, "ensure_studio_backend_path", lambda: None)
+
+    inference.load_chat_backend(
+        "fake-model",
+        hf_token = None,
+        max_seq_length = 2048,
+        load_in_4bit = True,
+        tensor_parallel = True,
+    )
+
+    assert calls[0]["tensor_parallel"] is True
+    assert calls[0]["mlx_distributed"] is True
diff --git a/unsloth_cli/tests/test_start.py b/unsloth_cli/tests/test_start.py
index affc24626e..065972b275 100644
--- a/unsloth_cli/tests/test_start.py
+++ b/unsloth_cli/tests/test_start.py
@@ -92,6 +92,7 @@ def test_claude_flags_detected_when_version_not_first_token(monkeypatch):
 
 def test_install_agent_prompts_then_installs(monkeypatch):
     # TTY + yes: run the documented install command, then re-resolve the now-present binary.
+    monkeypatch.setattr(start.os, "name", "posix")
     monkeypatch.setattr(start.sys, "stdin", SimpleNamespace(isatty = lambda: True))
     monkeypatch.setattr(start.typer, "confirm", lambda *a, **k: True)
     ran = []
@@ -108,6 +109,127 @@ def test_install_agent_prompts_then_installs(monkeypatch):
     assert ran == [["/bin/sh", "-c", "npm install -g @openai/codex"]]
 
 
+def test_install_agent_uses_powershell_on_windows(monkeypatch):
+    monkeypatch.setattr(start.os, "name", "nt")
+    monkeypatch.setattr(start.sys, "stdin", SimpleNamespace(isatty = lambda: True))
+    monkeypatch.setattr(start.typer, "confirm", lambda *a, **k: True)
+    ran = []
+    monkeypatch.setattr(
+        start.subprocess,
+        "run",
+        lambda command, *a, **k: ran.append(command) or SimpleNamespace(returncode = 0),
+    )
+    monkeypatch.setattr(start.shutil, "which", lambda _: r"C:\Users\samle\bin\hermes.exe")
+
+    install_hint = "& ([scriptblock]::Create((irm https://x/install.ps1))) -SkipSetup"
+    executable = start._install_agent("hermes", install_hint)
+
+    assert executable == r"C:\Users\samle\bin\hermes.exe"
+    assert ran == [["powershell", "-NoProfile", "-Command", install_hint]]
+
+
+def test_install_agent_warns_and_names_remote_source(monkeypatch, capsys):
+    # Before the confirm, a remote installer must name the URL it fetches so the
+    # user consents to a specific source rather than blindly accepting.
+    monkeypatch.setattr(start.os, "name", "nt")
+    monkeypatch.setattr(start.sys, "stdin", SimpleNamespace(isatty = lambda: True))
+    monkeypatch.setattr(start.typer, "confirm", lambda *a, **k: False)  # decline: nothing runs
+    hint = "& ([scriptblock]::Create((irm https://hermes-agent.nousresearch.com/install.ps1))) -SkipSetup"
+    assert start._install_agent("hermes", hint) is None
+    err = capsys.readouterr().err
+    assert "https://hermes-agent.nousresearch.com/install.ps1" in err
+    assert "download and RUN" in err
+    assert "signature or hash" in err
+
+
+def test_install_agent_warns_for_package_installer(monkeypatch, capsys):
+    # An npm-style installer has no URL to fetch, but still runs with the user's
+    # privileges, so the warning names the command instead.
+    monkeypatch.setattr(start.os, "name", "posix")
+    monkeypatch.setattr(start.sys, "stdin", SimpleNamespace(isatty = lambda: True))
+    monkeypatch.setattr(start.typer, "confirm", lambda *a, **k: False)
+    assert start._install_agent("codex", "npm install -g @openai/codex") is None
+    err = capsys.readouterr().err
+    assert "npm install -g @openai/codex" in err
+    assert "with your privileges" in err
+
+
+def test_hermes_install_hint_is_windows_native_on_windows(monkeypatch):
+    monkeypatch.setattr(start.os, "name", "nt")
+
+    # Scriptblock form so `-SkipSetup` reaches the installer and the interactive
+    # setup wizard is skipped during the unattended `unsloth start hermes` run.
+    assert start._hermes_install_hint() == (
+        "& ([scriptblock]::Create((irm https://hermes-agent.nousresearch.com/install.ps1)))"
+        " -SkipSetup"
+    )
+
+
+def test_hermes_install_hint_is_bash_on_posix(monkeypatch):
+    monkeypatch.setattr(start.os, "name", "posix")
+
+    # `bash -s -- --skip-setup` forwards the skip flag to the piped installer.
+    assert start._hermes_install_hint() == (
+        "curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent"
+        "/main/scripts/install.sh | bash -s -- --skip-setup"
+    )
+
+
+def test_refresh_windows_path_noop_off_windows(monkeypatch):
+    monkeypatch.setattr(start.os, "name", "posix")
+    before = os.environ.get("PATH", "")
+    monkeypatch.setenv("PATH", before)
+    start._refresh_windows_path()
+    assert os.environ.get("PATH", "") == before
+
+
+def test_refresh_windows_path_merges_registry_hives(monkeypatch):
+    # Fake Windows registry PATH values written after this process started.
+    hkcu, hklm = object(), object()
+    reg = {
+        (hkcu, "Environment"): r"C:\existing;C:\Users\me\hermes\bin",
+        (
+            hklm,
+            r"SYSTEM\CurrentControlSet\Control\Session Manager\Environment",
+        ): r"C:\Windows\System32",
+    }
+
+    class _Key:
+        def __init__(self, value):
+            self._value = value
+
+        def __enter__(self):
+            return self
+
+        def __exit__(self, *a):
+            return False
+
+    def open_key(root, sub):
+        if (root, sub) in reg:
+            return _Key(reg[(root, sub)])
+        raise OSError("missing hive")
+
+    fake_winreg = SimpleNamespace(
+        HKEY_CURRENT_USER = hkcu,
+        HKEY_LOCAL_MACHINE = hklm,
+        OpenKey = open_key,
+        QueryValueEx = lambda key, name: (key._value, 1),
+    )
+    monkeypatch.setattr(start.os, "name", "nt")
+    monkeypatch.setattr(start.os, "pathsep", ";")
+    monkeypatch.setitem(sys.modules, "winreg", fake_winreg)
+    monkeypatch.setenv("PATH", r"C:\custom;C:\existing")
+
+    start._refresh_windows_path()
+
+    assert os.environ["PATH"].split(";") == [
+        r"C:\custom",
+        r"C:\existing",
+        r"C:\Users\me\hermes\bin",
+        r"C:\Windows\System32",
+    ]
+
+
 def test_install_agent_declined_returns_none(monkeypatch):
     # TTY + no: never runs anything; caller falls back to the print-hint failure.
     monkeypatch.setattr(start.sys, "stdin", SimpleNamespace(isatty = lambda: True))
@@ -361,6 +483,189 @@ def test_connect_codex_no_launch(fake_studio, tmp_path):
     assert (home / "unsloth_api.config.toml").exists()
 
 
+def test_connect_codex_matches_requested_model_case_insensitively(fake_studio, tmp_path):
+    result = CliRunner().invoke(
+        start.start_app,
+        [
+            "codex",
+            "--no-launch",
+            "--model",
+            "unsloth/gemma-4-26b-a4b-it-gguf",
+        ],
+    )
+    assert result.exit_code == 0, result.output
+    home = tmp_path / "agents" / "codex"
+    profile = _parse_toml((home / "unsloth_api.config.toml").read_text())
+    assert profile["model"] == MODEL["id"]
+
+
+def test_resolve_model_matches_loaded_canonical_case_after_load(monkeypatch):
+    calls = []
+    state = {"loaded": False}
+
+    def http_json(
+        method,
+        url,
+        token,
+        payload = None,
+        timeout = 30,
+        error = None,
+    ):
+        calls.append((method, url, payload))
+        if url.endswith("/v1/models"):
+            return {
+                "data": [
+                    {
+                        "id": "unsloth/gemma-4-E2B-it-GGUF" if state["loaded"] else "other/model",
+                        "context_length": 131072,
+                    }
+                ]
+            }
+        if url.endswith("/api/inference/load"):
+            state["loaded"] = True
+            return {"model": "unsloth/gemma-4-E2B-it-GGUF"}
+        raise AssertionError(f"unexpected request: {method} {url}")
+
+    monkeypatch.setattr(start, "_http_json", http_json)
+
+    entry = start._resolve_model(
+        BASE,
+        "sk-test",
+        "unsloth/gemma-4-e2b-it-gguf",
+        start.LoadOptions(gguf_variant = "UD-Q4_K_XL"),
+    )
+
+    assert entry["id"] == "unsloth/gemma-4-E2B-it-GGUF"
+    assert any(c[1].endswith("/api/inference/load") for c in calls)
+
+
+def test_resolve_model_loads_when_catalog_hit_is_not_loaded(monkeypatch):
+    # A cached-but-unloaded catalog entry (loaded == False) that only case-differs must
+    # not be treated as ready; the load endpoint must still be called so the requested
+    # model becomes resident instead of the agent preflighting a different backend.
+    calls = []
+    state = {"loaded": False}
+
+    def http_json(
+        method,
+        url,
+        token,
+        payload = None,
+        timeout = 30,
+        error = None,
+    ):
+        calls.append((method, url))
+        if url.endswith("/v1/models"):
+            return {
+                "data": [
+                    {
+                        "id": "unsloth/Gemma-4-GGUF",
+                        "loaded": state["loaded"],
+                        "context_length": 131072,
+                    }
+                ]
+            }
+        if url.endswith("/api/inference/load"):
+            state["loaded"] = True
+            return {"model": "unsloth/Gemma-4-GGUF"}
+        raise AssertionError(f"unexpected request: {method} {url}")
+
+    monkeypatch.setattr(start, "_http_json", http_json)
+
+    entry = start._resolve_model(BASE, "sk-test", "unsloth/gemma-4-gguf")
+
+    assert entry["id"] == "unsloth/Gemma-4-GGUF"
+    assert any(u.endswith("/api/inference/load") for _, u in calls)
+
+
+def test_resolve_model_attaches_to_loaded_catalog_hit_without_reload(monkeypatch):
+    # The mirror case: a loaded entry (loaded == True) that case-matches attaches with
+    # no /api/inference/load call.
+    calls = []
+
+    def http_json(
+        method,
+        url,
+        token,
+        payload = None,
+        timeout = 30,
+        error = None,
+    ):
+        calls.append((method, url))
+        if url.endswith("/v1/models"):
+            return {
+                "data": [{"id": "unsloth/Gemma-4-GGUF", "loaded": True, "context_length": 131072}]
+            }
+        raise AssertionError(f"unexpected request: {method} {url}")
+
+    monkeypatch.setattr(start, "_http_json", http_json)
+
+    entry = start._resolve_model(BASE, "sk-test", "unsloth/gemma-4-gguf")
+
+    assert entry["id"] == "unsloth/Gemma-4-GGUF"
+    assert not any(u.endswith("/api/inference/load") for _, u in calls)
+
+
+def test_resolve_model_remote_studio_does_not_casefold_attach(monkeypatch):
+    # Against a remote Studio the local existence probe cannot see server-side paths,
+    # so a case-variant loaded id must NOT attach without a load: it could be a distinct
+    # server-side path on a case-sensitive host. The load endpoint resolves the request.
+    calls = []
+    state = {"loaded": False}
+
+    def http_json(
+        method,
+        url,
+        token,
+        payload = None,
+        timeout = 30,
+        error = None,
+    ):
+        calls.append((method, url))
+        if url.endswith("/v1/models"):
+            return {
+                "data": [{"id": "unsloth/Gemma-4-GGUF", "loaded": True, "context_length": 131072}]
+            }
+        if url.endswith("/api/inference/load"):
+            state["loaded"] = True
+            return {"model": "unsloth/Gemma-4-GGUF"}
+        raise AssertionError(f"unexpected request: {method} {url}")
+
+    monkeypatch.setattr(start, "_http_json", http_json)
+
+    entry = start._resolve_model("http://10.0.0.5:8888", "sk-test", "unsloth/gemma-4-gguf")
+
+    # The load endpoint was consulted (no casefold shortcut), and we still attach to the
+    # server's canonical id it reports back.
+    assert entry["id"] == "unsloth/Gemma-4-GGUF"
+    assert any(u.endswith("/api/inference/load") for _, u in calls)
+
+
+def test_model_id_matching_does_not_casefold_local_paths(tmp_path):
+    existing_local = tmp_path / "Org" / "Foo"
+    existing_local.mkdir(parents = True)
+
+    assert start._model_id_matches("Org/Foo", "org/foo")
+    assert not start._model_id_matches(str(existing_local), str(existing_local).lower())
+    assert not start._model_id_matches("./Models/Foo", "./models/foo")
+    assert not start._model_id_matches(r".\Models\Foo", r".\models\foo")
+    # A server-side relative path (extra path segments) is not a hub id even when it
+    # does not exist on the CLI host, so it must not casefold-match a differently
+    # cased path on a case-sensitive server filesystem.
+    assert not start._is_hub_model_id("models/Llama/Foo.gguf")
+    assert not start._model_id_matches("models/Llama/Foo.gguf", "models/llama/foo.gguf")
+    # A genuine two-segment hub id still matches case-insensitively.
+    assert start._is_hub_model_id("unsloth/Gemma-3-4b-it-GGUF")
+    assert start._model_id_matches("unsloth/Gemma-3-4b-it-GGUF", "unsloth/gemma-3-4b-it-gguf")
+    # Casefolding is gated to loopback studios (allow_casefold). With it disabled (a
+    # remote studio, where a two-segment string could be a server-side path), even a
+    # genuine hub-id case variant must not match, so the load endpoint resolves it.
+    assert not start._model_id_matches(
+        "unsloth/Gemma-3-4b-it-GGUF", "unsloth/gemma-3-4b-it-gguf", allow_casefold = False
+    )
+    assert start._model_id_matches("unsloth/Foo", "unsloth/Foo", allow_casefold = False)
+
+
 def test_connect_codex_launch_uses_ephemeral_home(fake_studio, monkeypatch):
     # Launch mode writes config to a throwaway temp CODEX_HOME and removes it after
     # the agent exits; the user's real ~/.codex is never the target.
@@ -435,15 +740,10 @@ def test_opencode_inline_config_beats_project_config(fake_studio):
     # permissions) ride in OPENCODE_CONFIG_CONTENT, which outranks project config.
     result = CliRunner().invoke(start.start_app, ["opencode", "--no-launch", "--yolo"])
     assert result.exit_code == 0, result.output
-    content_line = next(
-        ln for ln in result.output.splitlines() if ln.startswith("export OPENCODE_CONFIG_CONTENT=")
-    )
-    inline = json.loads(
-        shlex.split(content_line.removeprefix("export OPENCODE_CONFIG_CONTENT="))[0]
-    )
-    assert inline["model"] == f"unsloth/{MODEL['id']}"
+    inline = _opencode_inline_config(result.output)
+    assert inline["model"] == f"{start._OPENCODE_PROVIDER}/{MODEL['id']}"
     assert inline["permission"] == {"edit": "allow", "bash": "allow", "webfetch": "allow"}
-    assert "sk-unsloth" not in content_line  # key stays in the private file
+    assert "sk-unsloth" not in result.output  # key stays in the private file, not the env
 
 
 def test_opencode_inline_config_omits_permission_without_yolo(fake_studio):
@@ -452,13 +752,8 @@ def test_opencode_inline_config_omits_permission_without_yolo(fake_studio):
     # user's project rules; clearing our own config is the fix, and the inline pins the model.
     result = CliRunner().invoke(start.start_app, ["opencode", "--no-launch"])
     assert result.exit_code == 0, result.output
-    content_line = next(
-        ln for ln in result.output.splitlines() if ln.startswith("export OPENCODE_CONFIG_CONTENT=")
-    )
-    inline = json.loads(
-        shlex.split(content_line.removeprefix("export OPENCODE_CONFIG_CONTENT="))[0]
-    )
-    assert inline["model"] == f"unsloth/{MODEL['id']}"
+    inline = _opencode_inline_config(result.output)
+    assert inline["model"] == f"{start._OPENCODE_PROVIDER}/{MODEL['id']}"
     assert "permission" not in inline
 
 
@@ -1365,10 +1660,34 @@ def test_connect_openclaw_no_launch(fake_studio, tmp_path):
     config = json.loads(config_path.read_text())
     assert config["models"]["providers"]["unsloth"]["apiKey"] == "sk-unsloth-feedfacefeedface"
     assert config["agents"]["defaults"]["model"]["primary"] == f"unsloth/{MODEL['id']}"
+    assert _launch_command(result.output) == ["openclaw", "tui", "--local"]
     # OpenAI /v1/chat/completions works on either backend — no GGUF gate.
     assert not any(c[1].endswith("/api/inference/status") for c in fake_studio)
 
 
+def test_connect_openclaw_no_launch_keeps_explicit_subcommand(fake_studio):
+    result = CliRunner().invoke(start.start_app, ["openclaw", "--no-launch", "crestodian"])
+    assert result.exit_code == 0, result.output
+    assert _launch_command(result.output) == ["openclaw", "crestodian"]
+
+
+def test_connect_openclaw_no_launch_passes_global_flags_through(fake_studio):
+    # OpenClaw globals (openclaw [--dev] [--profile <name>] <command>) precede the
+    # command, and tui does not accept them, so any passthrough args must be forwarded
+    # verbatim rather than rewritten into `openclaw tui --local <globals>`.
+    result = CliRunner().invoke(start.start_app, ["openclaw", "--no-launch", "--profile", "test"])
+    assert result.exit_code == 0, result.output
+    assert _launch_command(result.output) == ["openclaw", "--profile", "test"]
+
+
+def test_connect_openclaw_no_launch_keeps_explicit_tui(fake_studio):
+    result = CliRunner().invoke(
+        start.start_app, ["openclaw", "--no-launch", "tui", "--message", "hi"]
+    )
+    assert result.exit_code == 0, result.output
+    assert _launch_command(result.output) == ["openclaw", "tui", "--message", "hi"]
+
+
 # ── OpenCode (OpenAI /v1/chat/completions) ───────────────────────────
 
 
@@ -1376,14 +1695,17 @@ def test_write_opencode_config_fresh(tmp_path):
     path = tmp_path / "opencode.json"
     start.write_opencode_config(BASE, "sk-unsloth-abc", MODEL, path)
     config = json.loads(path.read_text())
-    provider = config["provider"]["unsloth"]
+    provider = config["provider"][start._OPENCODE_PROVIDER]
     assert provider["npm"] == "@ai-sdk/openai-compatible"
     assert provider["options"] == {"baseURL": f"{BASE}/v1", "apiKey": "sk-unsloth-abc"}
     # Context limit must be declared, or OpenCode treats it as 0 and disables compaction.
     assert provider["models"] == {
         MODEL["id"]: {"name": MODEL["id"], "limit": {"context": 131072, "output": 8192}}
     }
-    assert config["model"] == f"unsloth/{MODEL['id']}"
+    assert config["model"] == f"{start._OPENCODE_PROVIDER}/{MODEL['id']}"
+    # The overlay never writes disabled_providers; the dedicated provider id is one a
+    # user's disable list would not target, so nothing needs re-enabling.
+    assert "disabled_providers" not in config
     # Compaction buffer scaled to ~10% of the window (compact near 90%).
     assert config["compaction"] == {"auto": True, "reserved": 131072 // 10}
 
@@ -1391,18 +1713,98 @@ def test_write_opencode_config_fresh(tmp_path):
 def test_write_opencode_config_preserves_and_idempotent(tmp_path):
     path = tmp_path / "opencode.json"
     path.write_text(
-        json.dumps({"theme": "tokyonight", "provider": {"anthropic": {"name": "Anthropic"}}})
+        json.dumps(
+            {
+                "theme": "tokyonight",
+                "disabled_providers": ["ollama", "unsloth"],
+                "provider": {"anthropic": {"name": "Anthropic"}},
+            }
+        )
     )
     start.write_opencode_config(BASE, "sk-unsloth-abc", MODEL, path)
     config = json.loads(path.read_text())
     assert config["theme"] == "tokyonight"
+    # The overlay no longer edits disabled_providers; re-enabling unsloth is done in
+    # the inline layer, so an existing list here is preserved untouched.
+    assert config["disabled_providers"] == ["ollama", "unsloth"]
     assert config["provider"]["anthropic"]["name"] == "Anthropic"
-    assert config["provider"]["unsloth"]["options"]["baseURL"] == f"{BASE}/v1"
+    assert config["provider"][start._OPENCODE_PROVIDER]["options"]["baseURL"] == f"{BASE}/v1"
     before = path.read_text()
     start.write_opencode_config(BASE, "sk-unsloth-abc", MODEL, path)
     assert path.read_text() == before
 
 
+def test_write_opencode_config_keeps_foreign_disabled_providers(tmp_path):
+    # A user who disabled other providers (but not unsloth) must keep them disabled:
+    # the overlay must not rewrite disabled_providers, or those providers get silently
+    # re-enabled for the session.
+    path = tmp_path / "opencode.json"
+    path.write_text(json.dumps({"disabled_providers": ["openai", "gemini"]}))
+    start.write_opencode_config(BASE, "sk-unsloth-abc", MODEL, path)
+    config = json.loads(path.read_text())
+    assert config["disabled_providers"] == ["openai", "gemini"]
+
+
+def _opencode_inline_config(output: str) -> dict:
+    # --no-launch prints OPENCODE_CONFIG_CONTENT as a POSIX `export NAME=<shell-quoted>`
+    # line on Unix/WSL and a PowerShell `$env:NAME = "<escaped>"` line on native Windows;
+    # parse whichever the host emitted so the opencode tests are shell-agnostic.
+    name = "OPENCODE_CONFIG_CONTENT"
+    for raw in output.splitlines():
+        line = raw.strip()
+        if line.startswith(f"export {name}="):
+            return json.loads(shlex.split(line.removeprefix(f"export {name}="))[0])
+        prefix = f'$env:{name} = "'
+        if line.startswith(prefix) and line.endswith('"'):
+            escaped = line[len(prefix) : -1]
+            # Reverse _print_env's PowerShell escaping (backtick is the escape char).
+            value = escaped.replace("`$", "$").replace('`"', '"').replace("``", "`")
+            return json.loads(value)
+    raise AssertionError(f"{name} not found in:\n{output}")
+
+
+def test_opencode_inline_scopes_session_to_studio_provider(fake_studio):
+    # opencode filters even config-defined providers through enabled/disabled_providers,
+    # and a model pin does not bypass that gate. The inline overlay (session-only, highest
+    # layer, arrays replace) allowlists our provider and clears the denylist so the Studio
+    # model always loads regardless of the user's config, without reading or editing it.
+    result = CliRunner().invoke(start.start_app, ["opencode", "--no-launch"])
+    assert result.exit_code == 0, result.output
+    inline = _opencode_inline_config(result.output)
+    assert inline["enabled_providers"] == [start._OPENCODE_PROVIDER]
+    assert inline["disabled_providers"] == []
+    assert inline["model"] == f"{start._OPENCODE_PROVIDER}/{MODEL['id']}"
+    # small_model stays on the enabled provider too, so lightweight tasks do not resolve a
+    # filtered provider mid-session.
+    assert inline["small_model"] == f"{start._OPENCODE_PROVIDER}/{MODEL['id']}"
+
+
+def test_opencode_passthrough_flags_omit_model_flag(fake_studio):
+    # Any passthrough (top-level flags that may precede a subcommand, or a subcommand)
+    # is left untouched; --model is not injected. The model is pinned by the inline
+    # OPENCODE_CONFIG_CONTENT (highest layer) instead, so it is still forced.
+    result = CliRunner().invoke(start.start_app, ["opencode", "--no-launch", "--dir", "repo"])
+    assert result.exit_code == 0, result.output
+    command = _launch_command(result.output)
+    assert command == ["opencode", "--dir", "repo"]
+    assert "--model" not in command
+    assert (
+        _opencode_inline_config(result.output)["model"]
+        == f"{start._OPENCODE_PROVIDER}/{MODEL['id']}"
+    )
+
+
+def test_opencode_passthrough_subcommand_omits_model_flag(fake_studio):
+    # A passthrough subcommand (e.g. `serve`) takes the model from the pinned config;
+    # inserting --model before it would break opencode's arg parsing.
+    result = CliRunner().invoke(start.start_app, ["opencode", "--no-launch", "serve"])
+    assert result.exit_code == 0, result.output
+    command = _launch_command(result.output)
+    assert command[0] == "opencode"
+    assert command[1] == "serve"
+    assert "--model" not in command
+
+
 def test_connect_opencode_no_launch(fake_studio, tmp_path):
     result = CliRunner().invoke(start.start_app, ["opencode", "--no-launch"])
     assert result.exit_code == 0, result.output
@@ -1410,9 +1812,24 @@ def test_connect_opencode_no_launch(fake_studio, tmp_path):
     config_path = tmp_path / "agents" / "opencode" / "opencode.json"
     # OPENCODE_CONFIG overlay points at the session file, not the user's global config.
     _assert_env_set(result.output, "OPENCODE_CONFIG", str(config_path))
+    inline_config = _opencode_inline_config(result.output)
     config = json.loads(config_path.read_text())
-    assert config["provider"]["unsloth"]["options"]["apiKey"] == "sk-unsloth-feedfacefeedface"
-    assert config["model"] == f"unsloth/{MODEL['id']}"
+    provider = config["provider"][start._OPENCODE_PROVIDER]
+    assert provider["options"]["apiKey"] == "sk-unsloth-feedfacefeedface"
+    assert config["model"] == f"{start._OPENCODE_PROVIDER}/{MODEL['id']}"
+    # The session config file (a throwaway overlay, not the user's real config) does not
+    # carry provider filters; the session scoping rides in the inline env layer only.
+    assert "disabled_providers" not in config
+    assert "enabled_providers" not in config
+    assert inline_config == {
+        "model": f"{start._OPENCODE_PROVIDER}/{MODEL['id']}",
+        "small_model": f"{start._OPENCODE_PROVIDER}/{MODEL['id']}",
+        "enabled_providers": [start._OPENCODE_PROVIDER],
+        "disabled_providers": [],
+    }
+    # --no-launch prints an append-safe base command (no --model before a subcommand a
+    # driver may append); the model is forced by the inline pin above.
+    assert _launch_command(result.output) == ["opencode"]
     assert not any(c[1].endswith("/api/inference/status") for c in fake_studio)
 
 
@@ -1765,7 +2182,7 @@ def test_no_launch_rerun_clears_stale_opencode_yolo_permissions(fake_studio, tmp
     # revert to OpenCode's permissive "allow" default).
     assert config["permission"] == {"edit": "ask", "bash": "ask", "webfetch": "ask"}
     # The session provider survives the cleanup.
-    assert "unsloth" in config["provider"]
+    assert start._OPENCODE_PROVIDER in config["provider"]
 
 
 def test_no_launch_rerun_clears_stale_openclaw_yolo_state(fake_studio, tmp_path):
@@ -2131,3 +2548,145 @@ def test_session_config_no_launch_preserves_existing_state(fake_studio, tmp_path
     with start._session_config("codex", launch = False) as home2:
         assert home2 == home
         assert (home2 / "sessions" / "live.sqlite").read_text() == "state"
+
+
+# ── --persist: persist the agent session so it can be resumed ────────────────
+def test_session_config_persist_uses_stable_dir_and_survives(monkeypatch, tmp_path):
+    # --persist routes a launch to the stable Unsloth agents dir (the one --no-launch
+    # already uses) instead of a throwaway temp dir, and never wipes it on exit.
+    monkeypatch.setattr(start, "_agents_config_root", lambda: tmp_path / "agents")
+    with start._session_config("codex", launch = True, persist = True) as home:
+        assert home == tmp_path / "agents" / "codex"
+        (home / "marker").write_text("kept")
+    assert home.exists()
+    assert (home / "marker").read_text() == "kept"
+
+
+def test_session_config_default_launch_is_ephemeral():
+    # Default launch (no --persist) still uses a throwaway temp dir wiped on exit.
+    with start._session_config("codex", launch = True) as home:
+        assert home.exists()
+        assert "unsloth-codex-" in home.name
+    assert not home.exists()
+
+
+# The temp-dir agents: --persist points each one's home/state env at the stable dir;
+# without it, at an ephemeral temp path. opencode is handled separately (only its
+# config overlay is relocated; its session data was never in the temp dir).
+_RESUME_ENV_VAR = {
+    "codex": "CODEX_HOME",
+    "openclaw": "OPENCLAW_STATE_DIR",
+    "hermes": "HERMES_HOME",
+    "pi": "HOME",
+}
+
+
+def _capture_launch(monkeypatch, argv):
+    captured = {}
+
+    def run(
+        command,
+        env = None,
+        **kwargs,
+    ):
+        captured["command"] = command
+        captured["env"] = env
+        return SimpleNamespace(returncode = 0)
+
+    monkeypatch.setattr(start.subprocess, "run", run)
+    result = CliRunner().invoke(start.start_app, argv)
+    assert result.exit_code == 0, result.output
+    return captured
+
+
+@pytest.mark.parametrize("agent", sorted(_RESUME_ENV_VAR))
+def test_resume_persists_agent_home_to_stable_dir(agent, fake_studio, tmp_path, monkeypatch):
+    monkeypatch.setattr(start.shutil, "which", lambda _: f"/usr/local/bin/{agent}")
+    captured = _capture_launch(monkeypatch, [agent, "--persist"])
+    stable = tmp_path / "agents" / agent
+    assert captured["env"][_RESUME_ENV_VAR[agent]] == str(stable)
+    # The stable dir survives the agent exit, so the session can be resumed.
+    assert stable.exists()
+
+
+@pytest.mark.parametrize("agent", sorted(_RESUME_ENV_VAR))
+def test_default_launch_home_is_ephemeral(agent, fake_studio, tmp_path, monkeypatch):
+    monkeypatch.setattr(start.shutil, "which", lambda _: f"/usr/local/bin/{agent}")
+    captured = _capture_launch(monkeypatch, [agent])
+    home = captured["env"][_RESUME_ENV_VAR[agent]]
+    assert f"unsloth-{agent}-" in home
+    assert str(tmp_path / "agents") not in home
+
+
+def test_resume_opencode_config_in_stable_dir(fake_studio, tmp_path, monkeypatch):
+    # opencode's session data lives in ~/.local/share/opencode (never relocated), so
+    # resume already survives exit; --persist also stabilizes its config overlay dir.
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/opencode")
+    captured = _capture_launch(monkeypatch, ["opencode", "--persist"])
+    stable = tmp_path / "agents" / "opencode"
+    assert captured["env"]["OPENCODE_CONFIG"] == str(stable / "opencode.json")
+    assert stable.exists()
+
+
+def test_persist_bare_codex_launch_has_no_resume_token(fake_studio, monkeypatch):
+    # A bare `--persist` only persists the session dir; it must NOT auto-append a native
+    # resume token, or the very first launch (no session yet) would send codex down its
+    # no-session error path. The user resumes explicitly: `unsloth start codex --persist resume`.
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/codex")
+    captured = _capture_launch(monkeypatch, ["codex", "--persist"])
+    assert "resume" not in captured["command"]
+    # command[0] is the resolved executable path; assert the argv after it.
+    assert captured["command"][1:] == ["--oss", "--profile", start._CODEX_PROFILE]
+
+
+def test_persist_bare_opencode_launch_has_no_resume_token(fake_studio, monkeypatch):
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/opencode")
+    captured = _capture_launch(monkeypatch, ["opencode", "--persist"])
+    assert "--continue" not in captured["command"]
+    assert captured["command"][1:] == ["--model", f"{start._OPENCODE_PROVIDER}/{MODEL['id']}"]
+
+
+def test_persist_bare_claude_launch_has_no_resume_token(fake_studio, monkeypatch):
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/claude")
+    monkeypatch.setattr(start, "_claude_flags", lambda: [])
+    captured = _capture_launch(monkeypatch, ["claude", "--persist"])
+    assert "--continue" not in captured["command"]
+    assert captured["command"][1:] == ["--model", MODEL["id"]]
+
+
+def test_resume_with_passthrough_does_not_auto_append(fake_studio, monkeypatch):
+    # When the caller drives their own subcommand, --persist only persists the dir; it
+    # must not inject a resume token that would collide with the user's command.
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/codex")
+    captured = _capture_launch(monkeypatch, ["codex", "--persist", "exec", "hello"])
+    assert "resume" not in captured["command"]
+    assert captured["command"][-2:] == ["exec", "hello"]
+
+
+def test_default_launch_has_no_resume_token(fake_studio, monkeypatch):
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/codex")
+    captured = _capture_launch(monkeypatch, ["codex"])
+    assert "resume" not in captured["command"]
+
+
+def test_resume_persist_only_agents_have_no_resume_token(fake_studio, monkeypatch):
+    # openclaw/hermes persist their session dir but have no non-interactive resume
+    # selector, so --persist must not append a token; their own picker resumes.
+    for agent in ("openclaw", "hermes"):
+        monkeypatch.setattr(start.shutil, "which", lambda _, a = agent: f"/usr/local/bin/{a}")
+        captured = _capture_launch(monkeypatch, [agent, "--persist"])
+        assert "resume" not in captured["command"]
+        assert "--continue" not in captured["command"]
+
+
+def test_native_resume_flag_passes_through_unchanged(fake_studio, monkeypatch):
+    # The persistence flag is --persist, NOT --resume, so an agent's own
+    # `--resume <id>` (e.g. `unsloth start claude --resume <guid>`) still flows
+    # through to the agent verbatim and is not swallowed as a Studio option.
+    monkeypatch.setattr(start.shutil, "which", lambda _: "/usr/local/bin/claude")
+    monkeypatch.setattr(start, "_claude_flags", lambda: [])
+    captured = _capture_launch(monkeypatch, ["claude", "--resume", "some-session-guid"])
+    assert captured["command"][-2:] == ["--resume", "some-session-guid"]
+    # Studio never auto-appends its own resume token when the user drives resume.
+    assert captured["command"].count("--resume") == 1
+    assert "--continue" not in captured["command"]