[LLM:Bugfix] Fix head_dim auto-detection for gated attention models (Qwen3.5)

Discussed-in: Merge-Request 26812010 , URL: https://code.alibaba-inc.com/AliNN/AliNNPrivate/codereview/26812010
GitOrigin-RevId: 3ea07dc659d9f59e800160d13281e635119ce2bc

.pre-commit-hooks/clang-format-diff.sh

@@ -8,17 +8,7 @@ if ! command -v git-clang-format &> /dev/null; then
     exit 1
 fi
 
-# Collect staged files with matching extensions, excluding generated headers
-files=$(git diff --cached --name-only --diff-filter=ACM \
-    | grep -E '\.(cpp|c|h|hpp|cc|m|mm)$' \
-    | grep -v 'MNN_generated\.h$' \
-    || true)
-
-if [ -z "$files" ]; then
-    exit 0
-fi
-
-output=$(git clang-format --diff --staged --extensions cpp,c,h,hpp,cc,m,mm -- $files 2>&1)
+output=$(git clang-format --diff --staged --extensions cpp,c,h,hpp,cc,m,mm 2>&1)
 
 if [ "$output" = "no modified files to format" ] || \
    [ "$output" = "clang-format did not modify any files" ]; then
@@ -31,7 +21,7 @@ if echo "$output" | grep -q "^diff"; then
     echo "$output"
     echo ""
     echo "To fix, run:"
-    echo "  git diff --cached --name-only --diff-filter=ACM | grep -E '\.(cpp|c|h|hpp|cc|m|mm)$' | grep -v 'MNN_generated\.h$' | xargs git clang-format --staged --"
+    echo "  git clang-format --staged --extensions cpp,c,h,hpp,cc,m,mm"
     echo "  git add -u"
     exit 1
 fi

transformers/llm/export/utils/transformers.py

@@ -102,6 +102,10 @@ class Attention(torch.nn.Module):
         # where head_dim varies per layer type: sliding=256, full=512)
         if hasattr(self, 'q_proj') and self.q_proj is not None and not (hasattr(self, 'qkv_proj') and self.qkv_proj is not None):
             actual_head_dim = self.q_proj.out_features // self.num_heads
+            # If q_norm exists, use its weight size as authoritative head_dim
+            # (q_proj may output 2x head_dim for gated attention, e.g. Qwen3.5)
+            if hasattr(self, 'q_norm') and self.q_norm is not None and hasattr(self.q_norm, 'weight'):
+                actual_head_dim = self.q_norm.weight.shape[0]
             if actual_head_dim != self.head_dim:
                 self.head_dim = actual_head_dim
                 # Re-detect num_key_value_heads with new head_dim