Add NaN debug checks after L1 and L2 GEMM

vllm/nvfp4_cutedsl.py

@@ -452,6 +452,11 @@ class CuTeDSLMoERunner:
             global_scale_a=l1_gsa, global_scale_b=self._l1_gsb,
         )
         
+        # DEBUG: Check for NaN after L1 GEMM
+        if torch.isnan(l1_out).any():
+            nan_count = torch.isnan(l1_out).sum().item()
+            print(f"[CLAWMINE] NaN in L1 output! {nan_count}/{l1_out.numel()} values are NaN")
+        
         # Extract real token outputs from padded GEMM output
         l1_out_real = l1_out[padded_dst]
         
@@ -488,6 +493,11 @@ class CuTeDSLMoERunner:
             global_scale_a=l2_gsa, global_scale_b=self._l2_gsb,
         )
         
+        # DEBUG: Check for NaN after L2 GEMM
+        if torch.isnan(l2_out).any():
+            nan_count = torch.isnan(l2_out).sum().item()
+            print(f"[CLAWMINE] NaN in L2 output! {nan_count}/{l2_out.numel()} values are NaN")
+        
         l2_out_real = l2_out[padded_dst]
         
         # === Scatter -> final output ===