This is a great project and thanks for open-sourcing. I have one question about the process of re-activating a woker/instance. According to the current implementation (python/sglang/srt/model_executor/worker_pool_model_runner.py#L397), the worker's cuda_graph_runner will be destroyed when it is deactivated. But it seems that when re-activating the worker，its graph will not be re-created (worker_pool_model_runner.py#L303) ? I wonder whether I understand this correct.
Looking forward to the team's response. Thank you!
This is a great project and thanks for open-sourcing. I have one question about the process of re-activating a woker/instance. According to the current implementation (python/sglang/srt/model_executor/worker_pool_model_runner.py#L397), the worker's cuda_graph_runner will be destroyed when it is deactivated. But it seems that when re-activating the worker，its graph will not be re-created (worker_pool_model_runner.py#L303) ? I wonder whether I understand this correct.
Looking forward to the team's response. Thank you!