diff --git a/vllm/patches/deepseek_v4_attention.py b/vllm/patches/deepseek_v4_attention.py
index c8182531..446dc1c6 100644
--- a/vllm/patches/deepseek_v4_attention.py
+++ b/vllm/patches/deepseek_v4_attention.py
@@ -751,6 +751,13 @@ class DeepseekV4MultiHeadLatentAttentionWrapper(PluggableLayer):
                 o[num_decode_tokens:] = causal_prefill_attention(
                     q_prefill, kv_rope_prefill, self.scale,
                 )
+            # Debug: check attention output
+            import sys as _sys
+            _amax = o[num_decode_tokens:].amax().item()
+            _nan = torch.isnan(o[num_decode_tokens:]).any().item()
+            _std = o[num_decode_tokens:].float().std().item()
+            if _amax > 100 or _nan or _std < 0.001:
+                print(f"[BLACKWELL] PREFILL CHECK: amax={_amax:.4f} NaN={_nan} std={_std:.6f} cr={self.compress_ratio}", file=_sys.stderr, flush=True)
 
         # Write into the output buffer
         if self.n_local_heads < self.padded_heads: