diff --git a/vllm/patches/deepseek_v4.py b/vllm/patches/deepseek_v4.py
index 3955184e..77c0363f 100644
--- a/vllm/patches/deepseek_v4.py
+++ b/vllm/patches/deepseek_v4.py
@@ -2220,6 +2220,30 @@ class DeepseekV4ForCausalLM(nn.Module):
         else:
             print("[POST-LOAD] No NaN scale tensors found — scales are clean")
         
+        # Dump layer 0 attn keys: model state_dict vs checkpoint
+        if int(os.environ.get('NVFP4_DEBUG', '0')):
+            sd = self.state_dict()
+            layer0_keys = sorted(k for k in sd if 'layers.0.attn' in k or 'layers.0.self_attn' in k)
+            print("=== MODEL state_dict (layer 0 attn): ===")
+            for k in layer0_keys:
+                t = sd[k]
+                nz = (t != 0).any().item() if torch.is_tensor(t) else '?'
+                print(f"  {k} shape={tuple(t.shape)} dtype={t.dtype} any_nonzero={nz}")
+            
+            from safetensors import safe_open
+            import glob
+            ckpt_files = sorted(glob.glob(os.path.join('/model', '*.safetensors')))
+            print(f"\n=== CHECKPOINT files: {len(ckpt_files)} shards ===")
+            seen = set()
+            for f in ckpt_files[:5]:
+                with safe_open(f, framework='pt') as h:
+                    for k in h.keys():
+                        if ('layers.0.' in k) and ('attn' in k or 'self_attn' in k):
+                            if k not in seen:
+                                seen.add(k)
+                                t = h.get_tensor(k)
+                                print(f"  {k} shape={tuple(t.shape)} dtype={t.dtype}")
+        
         return loaded_params
 
     def get_expert_mapping(self) -> list[tuple[str, str, int, str]]: