DEBUG: add sync + shape prints to shared_expert L2 gsa copy

dsv4/layers/shared_expert.py

@@ -363,7 +363,13 @@ class Nvfp4SharedExpert:
         # Fused amax + quantize: zero CPU syncs.
         if getattr(self, '_use_runtime_gsa', False):
             from dsv4.ops.quantize import quantize_nvfp4_gpu_fused
+            if not intermediate.is_contiguous():
+                intermediate = intermediate.contiguous()
             x_fp4, x_sf, gsa_l2_gpu = quantize_nvfp4_gpu_fused(intermediate)
+            # DEBUG: verify no CUDA errors from quantize kernel
+            torch.cuda.synchronize()  # DEBUG: catch async errors
+            if VERBOSE >= 1:
+                print(f"  SE L2 gsa: gsa_gpu shape={tuple(gsa_l2_gpu.shape)} dtype={gsa_l2_gpu.dtype} dev={gsa_l2_gpu.device} _l2_gsa_buf shape={tuple(self._l2_gsa_buf.shape)} dev={self._l2_gsa_buf.device}", flush=True)
             self._l2_gsa_buf.copy_(gsa_l2_gpu[:1].reshape(1))  # GPU → GPU, no sync
         else:
             x_fp4, x_sf = quantize_activation_nvfp4(