[checkpoint] Wait for async saves during close - #4270
Open
taking-lying-flat wants to merge 1 commit into
Open
Conversation
meta-cla
Bot
added
the
CLA Signed
taking-lying-flat
marked this pull request as ready for review
August 21, 2026 19:18
taking-lying-flat
requested review from
fegin,
tianyu-l,
wconstab and
wwwjn
as code owners
August 21, 2026 19:18
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
2 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
Wait for in-flight asynchronous checkpoint staging and saving before CheckpointManager releases its background resources.
Root cause
CheckpointManager tracks asynchronous work in staging_future and save_future. Normal checkpoint cycles wait for the previous save before starting the next one, but _close() only stopped the purge thread and closed the stager.
Shutdown paths do not necessarily start another save. In particular, dataloader exhaustion breaks out of training before the final-step save, while exceptions and explicit close calls also proceed directly to teardown. The last asynchronous upload could therefore still be running when close returned and the process group or process was torn down.
DefaultStager.close() drains its staging executor, but it does not await the checkpoint upload future or explicitly surface completion through the tracked futures. The close path now drains staging_future and save_future before releasing the purge and staging resources.