about DP scheduler

[guojinrong-nn]

I have some questions about DeepSeek’s DP (data parallelism) setup. Does the scheduler need to ensure that both the prefill and decode phases of the same request are assigned to the same DP rank? This is because the KV cache is stored on the rank where the prefill computation was performed. Is my understanding correct? If so, where is the implementation about this, I didn't found it.

[Huixxi]

A good question. I guess it may be implemented in vllm and sglang just reuse it innerly.

[zhyncs]

I guess it may be implemented in vllm and sglang just reuse it innerly

It's false. In the latest main, we have already removed the vllm dependency.

[guojinrong-nn]

Is there any details in the implementation?