[Feature] Support for Evicting Specific KV Cache to Save GPU Memory

[ChenlongDeng]

Checklist

• 1. If the issue you raised is not a feature but a question, please raise a discussion at https://github.com/sgl-project/sglang/discussions/new/choose Otherwise, it will be closed.
• 2. Please use English, otherwise it will be closed.

Motivation

Hi, congratulations on the amazing work!

I’d like to know if there is currently a feature that allows evicting specific parts of the KV cache (i.e., KV cache of some tokens) to save GPU memory. This capability is becoming increasingly important for many use cases involving KV cache compression, such as in methods like StreamingLLM and H2O.

I noticed that a similar issue was previously raised, and it was addressed with the introduction of DoubleSparse.(#1347, #1459 ) While DoubleSparse does reduce the computational cost of attention, it doesn’t seem to explicitly support operations for evicting specific parts of the KV cache from GPU memory.

I’m curious if such functionality is achievable within the current design of SGLang. If not, are there any plans to support this feature in the future?

Thank you!

StreamingLLM: https://arxiv.org/abs/2309.17453
H2O: https://arxiv.org/abs/2306.14048
DoubleSparsity: https://arxiv.org/abs/2408.07092

Related resources

No response

[github-actions]

This issue has been automatically closed due to inactivity. Please feel free to reopen it if needed.