Lightning-AI / Lightning-AI/pytorch-lightning

A new after on_save_checkpoint callback for AsyncCheckpointIO

Open
#19,414 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

checkpointing feature
Dominant language
Python
Stars
31.4k
Forks
3.8k
Avg merge
6d 7h
Merged PRs (30d)
6

Description

### Description & Motivation

We want to log_artifact, publish tags (e.g. with `MLFlowLogger`) only _after_ saving the checkpoint is complete (successfully). With `AsyncCheckpointIO` with does not seem to have a way to know when checkpoint saving is actually complete.

### Pitch

It would help to either add a new `on_after_save_checkpoint` to `CheckpointHooks` or to a new `AsyncCheckpointHooks`.
Then `AsyncCheckpointIO` can provide a way to register callbacks for the `Future`'s `add_done_callback`. When initializing the `Trainer` can iterate `Callback`s implementing the new method/hooks and register them with `Strategy` or `AsyncCheckpointIO`.

Regular `CheckpointIO` can also allow registration but will call the callbacks as soon as `save_checkpoint` is done.

### Alternatives

Currently the users are forced to copy-paste `AsyncCheckpointIO` and alter it to provide such hook.

### Additional context

I'm willing to contribute if there are no concerns with the proposed solution.

cc @lantiga @borda

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start by tracing how AsyncCheckpointIO saves checkpoints and how Trainer, Strategy, Callback, and CheckpointHooks interact. Review the Future.add_done_callback path and compare it with regular CheckpointIO completion. Done means callbacks can observe successful asynchronous completion while regular saves invoke them after saving finishes.

Written by the indexing model from the issue text.

Assessment

Tech stack
python, pytorch
Domain
machine-learning
Issue type
Feature
Difficulty
5/5
Estimated time
Over a week
Activity status
Stale
Clarity
Mostly clear
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.