CUDA-graph-compatible releasing and resuming KV cache and model weight memory #2630
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
# Conflicts: # python/sglang /srt/managers/tokenizer_manager.py
Co-authored-by: Lianmin Zheng <lianminzheng@gmail.com>
This reverts commit b03f558.
merrymercy
reviewed
Jan 13, 2025
| @@ -590,6 +596,7 @@ def init_memory_pool( | |||
| max_context_len=self.model_config.context_len + 4, | |||
| device=self.device, | |||
| use_records=False, | |||
| enable_memory_saver=self.server_args.enable_memory_saver, | |||
There was a problem hiding this comment.
See this comment #2630 (comment)
There was a problem hiding this comment.
Oops I misunderstood your question. I am worried maybe not, because it can happen that we do have some tensors inside req pool in the future that needs to be preserved across memory release. (The torch_memory_saver does not offload things to cpu, instead it just throws away the content, in order to be faster)
But anyway, today it seems to be OK (will check it), so I can update it if needed. Quickly skimmed it, worried maybe BaseTokenToKVPool.free_slots is one such tensor that hopes not to be released. Too tired now and not do any experiment though.
timethink
pushed a commit
to timethink/sglang
that referenced
this pull request
Mar 9, 2025
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Related: #2542 and #2583
Outdated Content
The test will fail because it uses
LD_PRELOADcurrently (to intercept and change logic of cudaMalloc and cudaFree). If the general logic looks good, I will further update this PR to handle this part (e.g. try to specify LD_PRELOAD automatically when creating the backend process.)How to execute it
Suppose this branch of SGLang is at
/path/to/sglang, then inside sglang's docker container, execute the following:Expected results are as follows. x is time, red color is memory consume. The low memory at the center is caused by temporarily release KV cache memory.
What's changed
Though the PR seems large, most are boilerplate.
Core:
with primary_memory_saver.region(): TokenToKVPool.k_buffers/v_buffers, ModelRunner.model, ReqToTokenPool.req_to_tokenprimary_memory_saver.pause()/.resume(): Atscheduler.py, Scheduler.release_gpu_occupation/resume_gpu_occupationOthers:
Checklist