Remove sync points in logits_processor + use median KL in tests#19646
Remove sync points in logits_processor + use median KL in tests#19646alisonshao wants to merge 5 commits intomainfrom
Conversation
alisonshao
requested review from
BBuf,
Edwardf0t1,
Fridge003,
HaiShaw,
Ying1123,
ch-wan,
hanming-lu,
hnyls2002,
ispobock,
merrymercy,
xiezhq-hermann and
yizhang2077
as code owners
Summary of ChangesHello, I'm Gemini Code Assist1! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed! This pull request reintroduces a previously reverted performance enhancement aimed at improving GPU utilization by eliminating unnecessary synchronization points. The re-landing is accompanied by a crucial fix to the KL divergence test, which was failing intermittently due to the performance changes. By adopting a median-based metric for KL divergence, the test now robustly handles occasional outliers, ensuring stability while retaining the performance benefits. Highlights
🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console. Changelog
Activity
Using Gemini Code AssistThe full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips. Invoking Gemini You can request assistance from Gemini at any point by creating a comment using either
Customization To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a Limitations & Feedback Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here. You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension. Footnotes
|
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request re-lands a patch to improve performance by removing GPU synchronization points and includes a fix for a flaky test that caused the original revert. The changes primarily involve using asynchronous data transfers (pin_memory=True, non_blocking=True), avoiding operations that cause implicit syncs (like .tolist() or not providing output_size to torch.repeat_interleave), and restructuring code to avoid CPU-GPU data dependencies. The fix for the flaky KL divergence test, which now uses the median instead of the mean to be robust against outliers, is a sound approach. The optimizations are well-implemented and should lead to better performance as intended. The code quality is high, and I have no specific comments for improvement.
|
/tag-and-rerun-ci |
|
thx for enhancing CI! #19639 |
|
@alisonshao Could you just make the test change only? The change re-land is done in #19639. |
|
sry no need for test only. can u have logits + test here? other stuff are covered in the linked pr above |
- Use pin_memory + non_blocking transfer for sample_indices and input_logprob_indices to avoid implicit GPU sync - Provide output_size to repeat_interleave to avoid internal sync - Use median instead of mean for KL divergence comparison to be robust against occasional single-prompt outliers
alisonshao
force-pushed
the
fix/sync-patch-logprob-fix
branch
from
a504a2e to
9cbf4c8
Compare
alisonshao
changed the title
|
/rerun-stage stage-b-test-large-1-gpu |
|
✅ Triggered |
|
@YazhiGao hi can you review |
3 participants
Summary
logits_processor.pyby usingpin_memory+non_blockingtransfer for tensor creation, and providingoutput_sizetorepeat_interleavecompare_kl_divergence(), making it robust against occasional single-prompt outliersFollowup to #19639 (which re-landed the other sync point removals) and #19581 (which reverted the original sync patch).
Motivation
The original sync patch (#19190) was reverted because async GPU transfers in
logits_processor.pyoccasionally produce single-prompt KL outliers (e.g., KL=0.34) that spike the mean above the 0.002 threshold, even though the vast majority of prompts have correct logprobs (~0.001 KL). Using median makes the test robust to these outliers while still catching real regressions.Test Results (10/10 passed on H200, full sync patch + median)
Runs 2 and 5 would have failed with the old mean-based metric. Median is consistently ~0.001, well below the 0.002 threshold.
Test plan
stage-b-test-large-1-gpushould pass