Fix NaN check: use os.environ gate instead of is_current_stream_capturing

torch.cuda.is_current_stream_capturing() returns bool, which breaks
Dynamo FX tracing (non-Tensor output). Switch to env var gate:
CLAWMINE_NAN_CHECK=1 enables NaN/Inf detection.

Dynamo evaluates os.environ at trace time — if the env var is not set,
the entire NaN check block is compiled away. Set it before first
inference to get NaN detection during prefill only.

vllm/patches/deepseek_v4.py

@@ -1,6 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
+import os
 import typing
 from collections.abc import Callable, Iterable
 from itertools import islice
@@ -1315,18 +1316,21 @@ class DeepseekV4Model(nn.Module):
                 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
+            # NaN detection — only during prefill. Disabled via env var during cudagraph.
+            # os.environ is evaluated at trace time by Dynamo, so the entire
+            # NaN check block is skipped during compilation.
+            if os.environ.get('CLAWMINE_NAN_CHECK', '0') == '1':
+                with torch.no_grad():
+                    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
+                    if layer_idx % 10 == 0:
+                        print(f"[CLAWMINE] Layer {layer_idx}: amax={hidden_states.amax().item():.4f} mean={hidden_states.mean().item():.6f}")
 
         # Stash pre-hc_head residual for the MTP draft (captured copy_).
         num_tokens = hidden_states.shape[0]