[Qwen3.5 bugfix] Add mm_input_embeds in piecewise cuda graph replay

python/sglang/srt/model_executor/piecewise_cuda_graph_runner.py

@@ -789,6 +789,11 @@ def replay(
                             if output.hidden_states is not None
                             else None
                         ),
+                        mm_input_embeds=(
+                            output.mm_input_embeds[: self.raw_num_tokens]
+                            if output.mm_input_embeds is not None
+                            else None
+                        ),
                     )
                 elif isinstance(output, EmbeddingPoolerOutput):
                     return output