[Question] Any plans to support pipeline RL to avoid ramp down time during weight update in sglang servers

[abhanshugpt]

Your Question

I am running slime for qwen3.5 397B model and the weight transfer is taking too long. I tried to profile the delay and it seems i am taking a lot of time in ramp down of requests in SGLang servers. Is it possible to avoid this ramp down during weight transfer similar to what has been done in pipeline RL paper: https://arxiv.org/pdf/2509.19128 for vllm. Let me know if its already supported or if there is an open PR to add this functionality.

What I've Tried

I have tried looking at documentation and FAQ but could not find the implementation or a direct way to use it in slime.

Environment (if relevant)

• slime version: v0.2.4

Additional Context

No response

Pre-submission Checklist

• I have read the CONTRIBUTING.md and understand the collaboration scope.
• I have read the documentation and FAQ and my question is not answered there.
• I have searched for existing issues and my question has not been asked before.

[zhuzilin]

I think that you can manually comment the flush_cache function in slime, which should be similar to pipeline RL.

[abhanshugpt]

I understand that we need to prevent calling flush_cache function, but before that we are calling pause_generation which abort all the current running requests, we need to fix this as well to just hold the current ongoing generation requests in a paused state till weight transfer is done and than resume generation after it.