Load state dicts to CPU #328
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
ibeltagy
reviewed
Oct 12, 2023
| load_fsdp_optim_state(self.fsdp_model, self.optim, optim_state_dict) | ||
|
|
||
| # Load other state. | ||
| try: | ||
| train_state_dict = torch.load(resource_path(load_path, "train.pt", local_cache=local_cache)) | ||
| train_state_dict = load_state_dict(load_path, "train.pt", local_cache=local_cache) |
There was a problem hiding this comment.
why no map_location="cpu" here?
There was a problem hiding this comment.
There's only one tiny tensor in the trainer state - the GPU RNG state. Which needs to go on GPU anyway.
ibeltagy
reviewed
Oct 12, 2023
| @@ -218,7 +218,13 @@ def save_state_dict( | |||
| upload(target_path, upload_target, save_overwrite=save_overwrite) | |||
There was a problem hiding this comment.
does this need a unit test?
There was a problem hiding this comment.
Our trainer unit tests are severely lacking at the moment, but I just tested this on spare MosaicML nodes are we recover the same loss exactly after these changes:
ibeltagy
approved these changes
Oct 12, 2023
|
@ibeltagy if this doesn't avoid OOM with our LUMI runs we will have to set |
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Outdated suggestions cannot be applied.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.
It turns out we can load (legacy) sharded and unsharded checkpoints to CPU via
torch.load()sinceFSDP.load_state_dict()will copy tensors to the right device anyway.This should save some GPU memory.