-
Notifications
You must be signed in to change notification settings - Fork 3.4k
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
Add lazy checkpoint loading for FSDP full-state checkpoints (Trainer) #18379
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
awaelchli
added
feature
Is an improvement or enhancement
checkpointing
Related to checkpointing
strategy: fsdp
Fully Sharded Data Parallel
and removed
pl
Generic label for PyTorch Lightning package
labels
Aug 24, 2023
awaelchli
force-pushed
the
feature/fsdp-lazy-load
branch
from
August 24, 2023 09:54
64218b8
to
a2c8780
Compare
for more information, see https://pre-commit.ci
awaelchli
changed the title
WIP: FSDP lazy load full state checkpoint
FSDP lazy load full state checkpoint
Aug 24, 2023
awaelchli
changed the title
FSDP lazy load full state checkpoint
Add lazy checkpoint loading for FSDP full-state checkpoints (Trainer)
Aug 24, 2023
awaelchli
requested review from
carmocca,
justusschock,
Borda and
williamFalcon
as code owners
August 24, 2023 10:03
carmocca
approved these changes
Aug 24, 2023
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This feature will likely be shortlived, but it's fine to add until the replacement is in
Borda
approved these changes
Aug 24, 2023
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Labels
checkpointing
Related to checkpointing
feature
Is an improvement or enhancement
pl
Generic label for PyTorch Lightning package
ready
PRs ready to be merged
strategy: fsdp
Fully Sharded Data Parallel
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?
Fixes #8043
Follows the same logic as added in #18150
Currently, loading a full checkpoint means replicating it in CPU memory for each process. This can cause CPU OOM on machines with limited CPU RAM. Instead, the approach of this PR is to lazy-load the checkpoint (no memory allocation) and on-the-fly load tensors as they are needed. This reduces the memory usage to only one weight tensor being stored in memory (per process) at a given time during
model.load_state_dict()
access.cc @Borda @awaelchli @carmocca