Adding logic for cleaning up FT checkpoints #1528
Open
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
bentherien
requested review from
tianyu-l,
fegin,
wwwjn and
wconstab
as code owners
meta-cla
bot
added
the
CLA Signed
bentherien
changed the title
tianyu-l
requested changes
Aug 5, 2025
torchtitan/config_manager.py
Outdated
| @@ -680,6 +680,11 @@ class FaultTolerance: | |||
| This is only used when "semi_sync_method" is set. | |||
| """ | |||
|
|
|||
| checkpoint_keep_latest_k: int = 0 | |||
There was a problem hiding this comment.
why can't you use checkpoint.keep_latest_k?
There was a problem hiding this comment.
When torchft is enabled there are two types of checkpoints:
- full checkpoint (also without torchft this can be enabled)
- per-replica checkpoint (specific for torchft per-step fault tolerance)
I believe this change should only affect 2, so it makes sense to keep it under the FaultTolerance dataclass
H-Huang
reviewed
Aug 5, 2025
torchtitan/components/checkpoint.py
Outdated
| @@ -112,14 +112,19 @@ def purge_thread(purge_queue: queue.Queue): | |||
| if isinstance(path, Terminate): | |||
| return | |||
| assert isinstance(path, str) | |||
| logger.info("Checkpointer is deleting %s.", path) | |||
|
|
|||
| if not 'ft-replica' in path: | |||
There was a problem hiding this comment.
Can "ft-replica" be a variable instead and use that across the checks? Also it looks like there is a mispelling for the folder name since its "replicat" currently lol, can you update that?
| @@ -641,6 +647,7 @@ def _ft_save(self, step: int) -> None: | |||
| self.save_future = self.dcp_save( | |||
| self.ft_states, checkpoint_id=checkpoint_id, async_mode=AsyncMode.ASYNC | |||
| ) | |||
| self._purge_stale_ft_checkpoints() | |||
There was a problem hiding this comment.
do we need to call this here? I thought the purge thread do the deletion
There was a problem hiding this comment.
Yes, but the directories to purge need to be added to the queue.
bentherien
force-pushed
the
ft_checkpoint_cleanup
branch
from
c19edff to
2818959
Compare
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.
When using semi-sync training, FT checkpoints can start to take up a considerable amount of storage and there is currectly no mechanism to clean them up. This PR adds:
For this initial PR, I made the decision to disable logging for this deletion as it creates too much output, but this is up for discussion.