perf: reduce DPO log-prob memory usage - #10720
Open
Tonystarkw12 wants to merge 1 commit into
Open
Conversation
Compute target-token log probabilities in sequence chunks instead of materializing a full-vocabulary log-softmax tensor. Preserve LD-DPO behavior and add numerical and gradient equivalence coverage. Co-Authored-By: Claude <noreply@anthropic.com>
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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?
Reduces peak memory usage during DPO log-probability computation.
Instead of materializing
logits.log_softmax(-1)for the full[batch,sequence,vocabulary]tensor, this computes target-token log probabilities in sequence chunksusing
gatherandlogsumexp. This preserves the same outputs and gradients while limiting temporary float32 allocations.LD-DPO keeps the existing implementation because it needs per-token log probabilities for its custom masking logic.
Motivation
The full-vocabulary float32 log-softmax tensor can cause out-of-memory failures during preference training, especially with long sequences and large vocabularies.
Chunking across the sequence dimension lowers temporary memory without changing model outputs.
Tests
get_batch_logpsruff checkandruff format --checkpassWANDB_DISABLED=true pytest -q --import-mode=importlib tests/train/test_dpo_trainer.py`passes```