diff --git a/dsv4/layers/shared_expert.py b/dsv4/layers/shared_expert.py
index 1cef8373..800889e4 100644
--- a/dsv4/layers/shared_expert.py
+++ b/dsv4/layers/shared_expert.py
@@ -374,6 +374,13 @@ class Nvfp4SharedExpert:
                 print(f"  SE L2: quantize_nvfp4_gpu_fused FAILED after sync: {e}", flush=True)
                 print(f"  intermediate: shape={tuple(intermediate.shape)} dtype={intermediate.dtype} dev={intermediate.device}", flush=True)
                 raise
+            # DEBUG: check gsa values before assignment
+            try:
+                gsa_first = gsa_l2_gpu[0].item()  # DEBUG: read value
+                print(f"  SE L2 gsa[0]={gsa_first:.6f} shape={tuple(gsa_l2_gpu.shape)} buf_shape={tuple(self._l2_gsa_buf.shape)}", flush=True)
+            except RuntimeError as e:
+                print(f"  SE L2: gsa[0].item() FAILED: {e}", flush=True)
+                raise
             # Copy first element of gsa to pre-allocated buffer.
             self._l2_gsa_buf[0] = gsa_l2_gpu[0]  # scalar GPU → GPU, no sync, graph-capturable
         else: