diff --git a/dsv4/layers/shared_expert.py b/dsv4/layers/shared_expert.py
index 800889e4..15267541 100644
--- a/dsv4/layers/shared_expert.py
+++ b/dsv4/layers/shared_expert.py
@@ -377,12 +377,13 @@ class Nvfp4SharedExpert:
             # 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)
+                print(f"  SE L2 gsa[0]={gsa_first:.6f} shape={tuple(gsa_l2_gpu.shape)} dev={gsa_l2_gpu.device} buf_dev={self._l2_gsa_buf.device} buf_shape={tuple(self._l2_gsa_buf.shape)}", flush=True)
+                # Try copy_ instead of scalar assign
+                self._l2_gsa_buf.copy_(gsa_l2_gpu[:1].contiguous())
+                print(f"  SE L2 gsa copy_ succeeded", flush=True)
             except RuntimeError as e:
-                print(f"  SE L2: gsa[0].item() FAILED: {e}", flush=True)
+                print(f"  SE L2: gsa assignment 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:
             x_fp4, x_sf = quantize_activation_nvfp4(
                 intermediate, self._l2_activation_global_scale