From 08cb60c6e1607f06888d1150dfcfddf95b3f135c Mon Sep 17 00:00:00 2001
From: shimmyshimmer <michaelhan2050@gmail.com>
Date: Thu, 23 Jul 2026 22:03:56 -0700
Subject: [PATCH] Studio autoload: retry the folder's next quant when a load
 itself fails

Fifth round of review follow-ups. A resolved quant that passed validation
could still fail /api/inference/load (corrupt file, llama.cpp startup
error); the cascade then abandoned the whole row because only validation
blocks recorded a skip key. The fallback now marks the failed quant
skipped and resolves the folder's next complete quant before moving on,
bounded by the existing attempt cap. Single-candidate rows resolve to
null once skipped, so the retry loop terminates.

Contract test and simulation added for the failure-then-retry path.
---
 .../src/features/chat/api/chat-adapter.ts     | 46 +++++++++++++++----
 tests/studio/test_model_picker_contracts.py   | 21 +++++++++
 2 files changed, 57 insertions(+), 10 deletions(-)

diff --git a/studio/frontend/src/features/chat/api/chat-adapter.ts b/studio/frontend/src/features/chat/api/chat-adapter.ts
index 095f74dc36..41d92f2df0 100644
--- a/studio/frontend/src/features/chat/api/chat-adapter.ts
+++ b/studio/frontend/src/features/chat/api/chat-adapter.ts
@@ -1568,8 +1568,12 @@ async function resolveLocalRowCandidate(
       return null;
     }
   }
+  const candidate = localRowToCandidate(row, isGguf ? rememberedVariant : null);
+  // Single-candidate rows resolve to null once their candidate is skipped,
+  // so the retry loop in the fallback terminates instead of re-attempting.
+  if (isSkippedCandidate?.(candidate)) return null;
   return {
-    candidate: localRowToCandidate(row, isGguf ? rememberedVariant : null),
+    candidate,
     sizeBytes: sizeOrUnknownBytes(row.size_bytes),
   };
 }
@@ -2262,20 +2266,42 @@ export async function autoLoadOnDeviceModel(): Promise<{
         continue;
       }
       const row = candidate.row;
-      const localCandidate = candidate.candidate;
+      let localCandidate: AutoLoadCandidate | null = candidate.candidate;
       if (isSeen(localCandidate.kind, row.load_id, row.id, row.path)) {
         continue;
       }
       markSeen(localCandidate.kind, row.load_id, row.id, row.path);
-      if (isSkippedAutoLoadCandidate(localCandidate)) {
-        continue;
-      }
-      try {
-        if (await loadAutoLoadCandidate(localCandidate)) {
-          return { loaded: true, blockedByTrustRemoteCode: false };
+      // Try the row's quants smallest-first: a failed LOAD (not just a
+      // blocked validation) marks that quant skipped and the folder's next
+      // complete quant is resolved and tried, so one corrupt file cannot
+      // abandon a folder that still holds a loadable quant. The attempt cap
+      // still bounds total /load calls; single-candidate rows resolve to
+      // null once skipped, terminating the loop.
+      while (localCandidate && loadAttempts < MAX_AUTO_LOAD_ATTEMPTS) {
+        if (!isSkippedAutoLoadCandidate(localCandidate)) {
+          try {
+            if (await loadAutoLoadCandidate(localCandidate)) {
+              return { loaded: true, blockedByTrustRemoteCode: false };
+            }
+          } catch {
+            hadNonTrustFailure = true;
+            skippedAutoLoadCandidates.add(
+              autoLoadCandidateKey(
+                localCandidate.kind,
+                localCandidate.id,
+                localCandidate.ggufVariant,
+              ),
+            );
+          }
+        }
+        try {
+          localCandidate =
+            (await resolveLocalRowCandidate(row, null, isSkippedAutoLoadCandidate))
+              ?.candidate ?? null;
+        } catch {
+          hadNonTrustFailure = true;
+          break;
         }
-      } catch {
-        hadNonTrustFailure = true;
       }
     }
 
diff --git a/tests/studio/test_model_picker_contracts.py b/tests/studio/test_model_picker_contracts.py
index f1c7037e1a..c5fb84e50f 100644
--- a/tests/studio/test_model_picker_contracts.py
+++ b/tests/studio/test_model_picker_contracts.py
@@ -708,3 +708,24 @@ def test_local_fallback_orders_by_resolved_quant_size():
         "const ggufGroup: FallbackCandidate[]"
     )
     assert "sizeBytes: resolved.sizeBytes" in auto_load
+
+
+def test_cascade_retries_next_quant_after_load_failure():
+    """A failed /api/inference/load (not just a blocked validation) must mark
+    that quant skipped and try the folder's next complete quant before the
+    row is abandoned; single-candidate rows resolve to null once skipped so
+    the retry loop terminates, and the attempt cap bounds total loads."""
+    src = _read("features/chat/api/chat-adapter.ts")
+    auto_load = src.split("async function autoLoadOnDeviceModel", 1)[1]
+    assert (
+        "while (localCandidate && loadAttempts < MAX_AUTO_LOAD_ATTEMPTS)" in auto_load
+    )
+    # The cascade catch records the failed quant, unlike the old generic flag.
+    local_loop = auto_load.split(
+        "while (localCandidate && loadAttempts < MAX_AUTO_LOAD_ATTEMPTS)", 1
+    )[1].split("\n    }", 1)[0]
+    assert "skippedAutoLoadCandidates.add(" in local_loop
+    # Termination guard: a skipped single candidate resolves to null.
+    resolve_fn = src.split("async function resolveLocalRowCandidate", 1)[1]
+    resolve_fn = resolve_fn.split("\nfunction ", 1)[0]
+    assert "if (isSkippedCandidate?.(candidate)) return null;" in resolve_fn