DEBUG: check SE input magnitude

dsv4/layers/shared_expert.py

@@ -425,6 +425,9 @@ class Nvfp4SharedExpert:
         """Actual implementation — called via custom autograd to be torch.compile-safe."""
         self._ensure_initialized()
 
+        # DEBUG: check input
+        print(f"  SE input: shape={tuple(hidden_states.shape)} |max|={hidden_states.abs().max().item():.6f} nan={torch.isnan(hidden_states).any().item()}", flush=True)
+
         if self._fused_swiglu:
             # P1: Fused L1 GEMM + SwiGLU + clamp in one kernel launch
             intermediate = self._run_l1_fused(hidden_states)