Make Trainer compatible with sharded checkpoints #17053
Merged
Conversation
This file contains 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
|
The documentation is not available anymore as the PR was closed or merged. |
LysandreJik
approved these changes
May 2, 2022
| """ | ||
| This is the same as | ||
| [`torch.nn.Module.load_state_dict`](https://pytorch.org/docs/stable/generated/torch.nn.Module.html?highlight=load_state_dict#torch.nn.Module.load_state_dict) | ||
| but for a sharded checkpoint. |
There was a problem hiding this comment.
I think it would be nice to have a documentation for sharded checkpoints so that users understand what they are. For example here having the sharded checkpoint redirect to a small blurb mentioning things like the following:
- What is it?
- Weights files that are split in multiple checkpoint
- Index showing how weights are linked
- Why is it important?
- Better to work with smaller files for memory
- Simpler to push to the hub
- How to work with it?
- Showing how to use
from_pretrainedandsave_pretrainedfor sharding - push to hub
- now trainer
- Showing how to use
Let me know if that's something that already exists, and if not I'm happy to help contribute it (or to contribute it altogether).
| @@ -327,6 +327,63 @@ def get_checkpoint_shard_files( | |||
| return cached_filenames, sharded_metadata | |||
|
|
|||
|
|
|||
| def load_sharded_checkpoint(model, folder, strict=True): | |||
There was a problem hiding this comment.
Should this be in the docs somewhere?
stevhliu
pushed a commit
to stevhliu/transformers
that referenced
this pull request
May 3, 2022
* Make Trainer compatible with sharded checkpoints * Add doc
elusenji
pushed a commit
to elusenji/transformers
that referenced
this pull request
Jun 12, 2022
* Make Trainer compatible with sharded checkpoints * Add doc
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.
What does this PR do?
The
Traineris currently incompatible with the new sharded checkpoint feature in two places:In both cases, the model state dict is loaded back inside the model but there is no model save file if the model was above the default size for sharding, resulting in errors (as was pointed out by #16976 ).
This PR addresses this by:
load_sharded_checkpointthat does the same thing asmodel.load_state_dictfor regular model files, but loads a sharded checkpoint (and errors in case of missing/unexpected keys whenstrict=True).A test is added to make sure resuming works from a sharded checkpoint.
Fixes #16976