Add NaN/Inf detection in DeepseekV4Model.forward layer loop

- Checks every layer during prefill (not during cudagraph capture)
- is_current_stream_capturing() gate prevents CPU-GPU syncs during capture
- Prints amax every 10 layers for magnitude tracking
- Breaks on first NaN/Inf to avoid wasting compute

vllm/patches/deepseek_v4.py

@@ -1309,12 +1309,24 @@ class DeepseekV4Model(nn.Module):
         hidden_states = hidden_states.unsqueeze(-2).repeat(1, self.hc_mult, 1)
         if self.use_mega_moe:
             input_ids = input_ids.to(torch.int64)
-        for layer in islice(self.layers, self.start_layer, self.end_layer):
+        for layer_idx, layer in enumerate(islice(self.layers, self.start_layer, self.end_layer)):
             hidden_states = layer(
                 hidden_states,
                 positions,
                 input_ids,
             )
+            # NaN detection (prefill only — no CPU-GPU syncs during cudagraph)
+            if not torch.cuda.is_current_stream_capturing():
+                if layer_idx % 10 == 0:
+                    print(f"[CLAWMINE] Layer {layer_idx}: amax={hidden_states.amax().item():.4f} mean={hidden_states.mean().item():.6f}")
+                if torch.isnan(hidden_states).any():
+                    nan_pct = torch.isnan(hidden_states).float().mean().item() * 100
+                    print(f"[CLAWMINE] NaN after layer {layer_idx}! {nan_pct:.2f}% NaN, amax={hidden_states.amax().item():.4f}")
+                    break
+                if torch.isinf(hidden_states).any():
+                    inf_pct = torch.isinf(hidden_states).float().mean().item() * 100
+                    print(f"[CLAWMINE] Inf after layer {layer_idx}! {inf_pct:.2f}% Inf, amax={hidden_states.amax().item():.4f}")
+                    break
 
         # Stash pre-hc_head residual for the MTP draft (captured copy_).
         num_tokens = hidden_states.shape[0]