When the `enforce_eager` in the config is set to `False`, the CUDA graph cannot be captured correctly.