fix: contiguous copy for SF byte view sanity check

deep_gemm/mega/__init__.py

@@ -351,7 +351,9 @@ def fp8_nvfp4_mega_moe(y: torch.Tensor,
         zero_pct = (sf == 0).float().mean().item() * 100
         if zero_pct > 50:
             print(f"[WARN] {name}: {zero_pct:.1f}% zeros in SF! Possible div-by-zero", flush=True)
-        sf_u8 = sf.view(torch.uint8)
+        # Need contiguous for byte view — make a contiguous copy just for the check
+        sf_c = sf.contiguous()
+        sf_u8 = sf_c.view(torch.uint8)
         nan_count = (sf_u8 == 0x7F).sum().item()
         inf_count = (sf_u8 == 0x7E).sum().item()
         if nan_count > 0 or inf_count > 0: