Add debug patch to print layer name mismatch

Dockerfile

@@ -65,6 +65,11 @@ RUN python3 /tmp/patch_swa_cache.py ${VLLM_SPARSE_SWA_DIR}/sparse_swa.py && rm /
 COPY vllm/patches/patch_indexer_cache.py /tmp/patch_indexer_cache.py
 RUN python3 /tmp/patch_indexer_cache.py ${VLLM_LAYERS_DIR2}/deepseek_v4_attention.py && rm /tmp/patch_indexer_cache.py
 
+# Debug: print layer name mismatch
+ARG VLLM_WORKER_DIR=/usr/local/lib/python3.12/dist-packages/vllm/v1/worker
+COPY vllm/patches/patch_debug_layers.py /tmp/patch_debug_layers.py
+RUN python3 /tmp/patch_debug_layers.py ${VLLM_WORKER_DIR}/gpu_model_runner.py && rm /tmp/patch_debug_layers.py
+
 # Register CuTeDSL kernel in vLLM's linear kernel selection
 ARG VLLM_LINEAR_DIR=/usr/local/lib/python3.12/dist-packages/vllm/model_executor/kernels/linear
 COPY vllm/patches/register_cutedsl_kernel.py /tmp/register_cutedsl_kernel.py

vllm/patches/patch_debug_layers.py

@@ -0,0 +1,39 @@
+#!/usr/bin/env python3
+"""Patch _allocate_kv_cache_tensors to print the layer name mismatch."""
+import sys
+
+def patch(path):
+    with open(path, 'r') as f:
+        content = f.read()
+
+    if "CLAWMINE_DEBUG_LAYERS" in content:
+        print("Already patched, skipping")
+        return
+
+    old = """        assert layer_names == set(kv_cache_raw_tensors.keys()), (
+            "Some layers are not correctly initialized"
+        )"""
+
+    new = """        # CLAWMINE_DEBUG_LAYERS: print mismatch instead of asserting
+        missing = layer_names - set(kv_cache_raw_tensors.keys())
+        extra = set(kv_cache_raw_tensors.keys()) - layer_names
+        if missing or extra:
+            print(f"CLAWMINE DEBUG: missing layers ({len(missing)}): {sorted(missing)[:20]}")
+            print(f"CLAWMINE DEBUG: extra layers ({len(extra)}): {sorted(extra)[:20]}")
+            print(f"CLAWMINE DEBUG: expected ({len(layer_names)}), got ({len(kv_cache_raw_tensors.keys())})")
+        assert layer_names == set(kv_cache_raw_tensors.keys()), (
+            "Some layers are not correctly initialized"
+        )"""
+
+    if old not in content:
+        print("ERROR: Could not find the code to patch")
+        sys.exit(1)
+
+    content = content.replace(old, new)
+
+    with open(path, 'w') as f:
+        f.write(content)
+    print("Patched gpu_model_runner.py for debug layer names")
+
+if __name__ == "__main__":
+    patch(sys.argv[1])