Skip to content

[BugFix] fix gpt-oss accuracy issue when enabling piecewise cuda graph#18013

Merged
ispobock merged 3 commits intosgl-project:mainfrom
zminglei:fix-gpt-oss-acc
Feb 1, 2026
Merged

[BugFix] fix gpt-oss accuracy issue when enabling piecewise cuda graph#18013
ispobock merged 3 commits intosgl-project:mainfrom
zminglei:fix-gpt-oss-acc

Conversation

@zminglei
Copy link
Copy Markdown
Collaborator

@zminglei zminglei commented Jan 31, 2026 •
edited
Loading

Motivation

fix #17839
fix gpt-oss accuracy issue when enabling piecewise cuda graph

Modifications

Accuracy Tests

Before:

python benchmark/gsm8k/bench_sglang.py --data-path /shared/public/data/gsm8k/test.jsonl
100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 200/200 [00:25<00:00,  7.73it/s]
Accuracy: 0.785
Invalid: 0.010
Latency: 25.943 s
Output throughput: 2590.943 token/s

After:

python benchmark/gsm8k/bench_sglang.py --data-path /shared/public/data/gsm8k/test.jsonl
100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 200/200 [00:25<00:00,  7.89it/s]
Accuracy: 0.845
Invalid: 0.015
Latency: 25.400 s
Output throughput: 2478.192 token/s

Benchmarking and Profiling

Checklist

Review Process

  1. Ping Merge Oncalls to start the PR flow. See the PR Merge Process.
  2. Get approvals from CODEOWNERS and other reviewers.
  3. Trigger CI tests with comments or contact authorized users to do so.
    • /tag-run-ci-label, /rerun-failed-ci, /tag-and-rerun-ci
  4. After green CI and required approvals, ask Merge Oncalls to merge.

@gemini-code-assist
Copy link
Copy Markdown
Contributor

gemini-code-assist Bot commented Jan 31, 2026

Summary of Changes

Hello @zminglei, 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 addresses and rectifies an accuracy degradation in GPT-OSS models that occurs when the piecewise CUDA graph optimization is active. The core change involves a strategic modification to the KV cache eviction logic, specifically preventing SWA eviction during the decode phase under these conditions. This ensures that CUDA graph memory addresses remain stable, thereby resolving the accuracy issue and enhancing the reliability of the model's performance with this optimization enabled.

Highlights

  • Accuracy Fix: Resolves a critical accuracy issue observed in GPT-OSS models when the piecewise CUDA graph feature is enabled, leading to improved model output quality.
  • KV Cache Management: Introduces a conditional skip for Sliding Window Attention (SWA) KV cache eviction during the decode phase. This prevents conflicts with memory addresses captured by CUDA graphs.
  • Performance Stability: Ensures the stability and correctness of the model's operation when advanced CUDA graph optimizations are active, preventing unintended side effects on memory management.
  • Benchmark Improvement: Demonstrates a significant accuracy improvement from 0.785 to 0.845 on the GSM8K benchmark after applying the fix.

🧠 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.

Using Gemini Code Assist

The 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 /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature Command Description
Code Review /gemini review Performs a code review for the current pull request in its current state.
Pull Request Summary /gemini summary Provides a summary of the current pull request in its current state.
Comment @gemini-code-assist Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help /gemini help Displays a list of available commands.

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 .gemini/ folder in the base of the repository. Detailed instructions can be found here.

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

  1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

gemini-code-assist[bot]
gemini-code-assist Bot reviewed Jan 31, 2026
View reviewed changes
Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Code Review

This pull request correctly fixes an accuracy issue with gpt-oss when piecewise CUDA graph is enabled by skipping SWA eviction during the decode phase. The change is logical and well-commented. I have one minor suggestion to improve efficiency by moving a function call out of a loop.

Comment thread python/sglang/srt/managers/schedule_batch.py Outdated
@zminglei zminglei marked this pull request as ready for review January 31, 2026 05:13
@gemini-code-assist
Copy link
Copy Markdown
Contributor

gemini-code-assist Bot commented Jan 31, 2026

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

hebiao064
hebiao064 reviewed Jan 31, 2026
View reviewed changes
Comment thread python/sglang/srt/managers/schedule_batch.py
hebiao064
hebiao064 reviewed Jan 31, 2026
View reviewed changes
Comment thread python/sglang/srt/managers/schedule_batch.py
@zminglei
Copy link
Copy Markdown
Collaborator Author

zminglei commented Jan 31, 2026

/tag-and-rerun-ci

@ispobock ispobock merged commit 38d275a into sgl-project:main Feb 1, 2026
171 of 186 checks passed
charlesHsuGG pushed a commit to charlesHsuGG/sglang that referenced this pull request Feb 2, 2026
@zminglei
[BugFix] fix gpt-oss accuracy issue when enabling piecewise cuda graph (
428547a 
sgl-project#18013)
sfiisf pushed a commit to sfiisf/sglang that referenced this pull request Feb 5, 2026
@zminglei @sfiisf
[BugFix] fix gpt-oss accuracy issue when enabling piecewise cuda graph (
0d69474 
sgl-project#18013)
Johnsonms pushed a commit to Johnsonms/sglang that referenced this pull request Feb 14, 2026
@zminglei @Johnsonms
[BugFix] fix gpt-oss accuracy issue when enabling piecewise cuda graph (
d5f8fc9 
sgl-project#18013)
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@hebiao064 hebiao064 hebiao064 left review comments

@ispobock ispobock ispobock approved these changes

@merrymercy merrymercy Awaiting requested review from merrymercy merrymercy is a code owner

@Ying1123 Ying1123 Awaiting requested review from Ying1123 Ying1123 is a code owner

@hnyls2002 hnyls2002 Awaiting requested review from hnyls2002 hnyls2002 is a code owner

@xiezhq-hermann xiezhq-hermann Awaiting requested review from xiezhq-hermann xiezhq-hermann is a code owner

+1 more reviewer

@gemini-code-assist gemini-code-assist[bot] gemini-code-assist[bot] left review comments

Reviewers whose approvals may not affect merge requirements

Assignees

No one assigned

Labels

run-ci

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

[Bug] enabling piecewise cuda graph would degrade the accuracy of gpt-oss-120b

3 participants

@zminglei @hebiao064 @ispobock