fix: remove .item() sync in scale reshape — use padded_scales.shape[0] instead

vllm/nvfp4_cutedsl.py

@@ -183,8 +183,9 @@ class CuTeDSLMoERunner:
         padded_scales[dst_rows, :K_sf] = x_sf
         
         # Apply swizzle to the whole padded tensor, return 2D for 2D-side scale_a
-        swizzled_flat = pad_and_swizzle_single(padded_scales)
-        return swizzled_flat.reshape(total_padded_rows.item(), -1)
+        # to_blocked preserves element count, so reshape to match padded shape
+        swizzled = pad_and_swizzle_single(padded_scales)
+        return swizzled.reshape(padded_scales.shape[0], -1)
 
     def run(self, hidden_states, topk_weights, topk_ids, expert_indices=None):
         num_tokens, hidden_size = hidden_states.shape