Repository navigation
fix(vllm): fail one request, not the engine, when the worker runs out of memory - #33
Merged
Merged
Conversation
… of memory An out-of-memory error in the worker's static harvest killed EngineCore, because the copy runs inside execute_model. - _harvest copies only each capturing request's own rows, one clone per request. A request's rows no longer keep its co-batched rows alive. On OOM it fails that request and the step goes on. - Collect, drain and the lens read-out take the rows out of the store first. A final call deregisters before it allocates, so an OOM there leaves no state behind. Chunk lists are freed as they are joined. - Under tensor parallelism the ranks agree on failure before the next collective, so a rank-local OOM raises on all ranks and does not hang. - Backend cleanups run shielded, and release runs even when collect raises, so a cancelled request does not leave rows in the worker. Co-authored-by: Cursor <cursoragent@cursor.com>
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
An OOM in the worker's static harvest copy killed EngineCore (seen on a 1×B200 DeepSeek-V4-Flash pod under concurrent lens requests).
Tests: ruff, pyright, CPU suite, and 30 vLLM GPU tests pass.
Made with Cursor