Optimize DPO log probability calculation by retaining necessary cache, saving up to 30GB of memory (#1968) #1969
Conversation
This file contains hidden or 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
SeungyounShin
commented
Aug 26, 2024
| reference_chosen_logps.append(reference_chosen_logp.cpu()) | ||
| reference_rejected_logps.append(reference_rejected_logp.cpu()) | ||
|
|
||
| # Unnecessary cache clearing to avoid OOM |
There was a problem hiding this comment.
Without this line, GPU memory usage can increase up to 45GB when running a LLaMA-8B model on an 8K context dataset:
Simply adding this line reduces and stabilizes memory consumption, keeping it under control:
There was a problem hiding this comment.
It seems that this cache occupies memory during training, leading to reduced available memory and, consequently, limiting the context size that can be used for training.
SeungyounShin
changed the title
SeungyounShin
changed the title
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
yxliu-TAMU
pushed a commit
to mincheolseong/ECEN743-GRPO-Project-Proposal
that referenced
this pull request
Apr 20, 2025
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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Outdated suggestions cannot be applied.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?
This PR addresses an issue where unnecessary cache clearing was performed during the log probability calculation in the training loop. The original code included calls to
torch.cuda.empty_cache()andself.accelerator.free_memory()to avoid out-of-memory (OOM) errors, but these were not necessary and may have introduced performance overhead. By removing these calls, this PR optimizes the memory management during training.Fixes # (issue)
Motivation and Context
This change improves the efficiency of the log probability calculation during training by eliminating redundant memory management operations. This can lead to better performance and more efficient use of GPU memory.
Before submitting
Who can review?
Anyone in the community is welcome to review this PR once the tests have passed. Feel free to tag members or contributors who may be interested in reviewing.