Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Initial commit for CheckpointManager #5678

Merged
merged 3 commits into from
Oct 10, 2023
Merged

Initial commit for CheckpointManager #5678

merged 3 commits into from
Oct 10, 2023

Conversation

jonb377
Copy link
Collaborator

@jonb377 jonb377 commented Oct 5, 2023

This adds the boilerplate code and documentation for CheckpointManager, a higher-level wrapper which interfaces with the torch.distributed.checkpoint APIs. CheckpointManager will provide some key functionality:

  • Asynchronous checkpointing
  • Autocheckpointing on preemption

@alanwaketan
Copy link
Collaborator

So this doesn't depend on #5630? That PR is a little bit stale nowadays. Should we prioritize to get that landed first? I saw @yeounoh has some unresolved comments.

@jonb377
Copy link
Collaborator Author

jonb377 commented Oct 5, 2023

So this doesn't depend on #5630? That PR is a little bit stale nowadays. Should we prioritize to get that landed first? I saw @yeounoh has some unresolved comments.

The functionality here doesn't depend on it. For distributed checkpointing, the main use case of that PR is to discover the master IP for the CPU process group. The stronger use case is going to be distributed data loading.

I'll sync with Yeounoh if we're OK to land it now, otherwise I can convert to draft until it's needed.

Copy link
Collaborator

@alanwaketan alanwaketan left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM. Please make sure this is up to date with the design doc.

Comment on lines +72 to +78
async_queue_size: The size of the execution queue which processes async
checkpoints. This should be a small value to ensure training doesn't
get too far ahead of the last finished checkpoint, but increasing
the value to 2 can unblock training when there are transient
network issues which slow down the active checkpoint.
Default: 1, which only allows a single async checkpoint to be
pending at a time.
Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@jonb377 jonb377 merged commit 4524543 into master Oct 10, 2023
18 checks passed
@jonb377 jonb377 deleted the jonbolin/chkpt-manager branch October 10, 2023 01:14
qihqi pushed a commit that referenced this pull request Oct 10, 2023
* Initial commit for CheckpointManager

* Update documentation for async

* Fix typo
zpcore pushed a commit that referenced this pull request Oct 19, 2023
* Initial commit for CheckpointManager

* Update documentation for async

* Fix typo
ghpvnist pushed a commit to ghpvnist/xla that referenced this pull request Oct 31, 2023
* Initial commit for CheckpointManager

* Update documentation for async

* Fix typo
mbzomowski pushed a commit to mbzomowski-test-org/xla that referenced this pull request Nov 16, 2023
* Initial commit for CheckpointManager

* Update documentation for async

* Fix typo
chunnienc pushed a commit to chunnienc/xla that referenced this pull request Dec 14, 2023
* Initial commit for CheckpointManager

* Update documentation for async

* Fix typo
golechwierowicz pushed a commit that referenced this pull request Jan 12, 2024
* Initial commit for CheckpointManager

* Update documentation for async

* Fix typo
bhavya01 pushed a commit that referenced this pull request Apr 22, 2024
* Initial commit for CheckpointManager

* Update documentation for async

* Fix typo
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants