[AMD] [DO NOT MERGE] [GLM-5 Day 0] Add GLM-5 nightly test#18911
[AMD] [DO NOT MERGE] [GLM-5 Day 0] Add GLM-5 nightly test#18911michaelzhang-ai wants to merge 13 commits intosgl-project:mainfrom
Conversation
…P/ROCm support in RotaryEmbedding class
…ROCm 7.2; refactor forward_hip method in RotaryEmbedding class to accept *args/**kwargs for better subclass compatibility.
- Introduced new jobs for 8-GPU GLM-5 accuracy tests in both AMD and MI35x workflows. - Updated workflow triggers to include pull requests for the main branch. - Added new test scripts for GLM-5 evaluation on MI325 and MI35x platforms. - Enhanced job definitions to install necessary dependencies and run accuracy benchmarks. This update improves CI coverage for GLM-5 models, ensuring robust testing across different hardware configurations.
Summary of ChangesHello @michaelzhang-ai, 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 significantly enhances AMD GPU support by introducing dedicated nightly accuracy tests for the GLM-5 large language model across different AMD MI series GPUs. It ensures that the model's performance on these platforms is consistently monitored and validated, and addresses a compatibility issue in the rotary embedding layer to facilitate proper execution on AMD hardware. 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
Ignored Files
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 adds nightly CI tests for GLM-5 on AMD GPUs. The changes include a necessary modification to rotary_embedding.py to support AMD GPUs and two new test files for different AMD hardware. The new test files contain a significant amount of duplicated code. To improve maintainability, I suggest refactoring the common logic into a shared utility module or a base test class in a follow-up PR. I've also suggested a small improvement to the ModelConfig dataclass in the new test files to use a more idiomatic approach for mutable default values.
…5 and refactor ModelConfig dataclass initialization
…e separate HIP implementation and consolidating the import logic for rotary_embedding.
…dling, ensuring correct rotary_embedding usage for each platform.
michaelzhang-ai
changed the title
michaelzhang-ai
requested review from
Fridge003,
Kangyan-Zhou,
ispobock and
merrymercy
as code owners
michaelzhang-ai
changed the title
…in test_gsm8k_eval_amd.py
|
Nightly 720 pass: https://github.com/sgl-project/sglang/actions/runs/22167228133 |
1 participant
Motivation
Add nightly CI accuracy tests for GLM-5 on AMD MI325/MI300X and MI35x runners.
GLM-5 is a 744B parameter (40B active) MoE model that uses DeepSeek Sparse Attention (DSA/NSA), sharing the same architecture family as DeepSeek-V3.2.
Nightly 720 pass: https://github.com/sgl-project/sglang/actions/runs/22167228133
Nightly pass: https://github.com/sgl-project/sglang/actions/runs/22167228136
Please help review. Thanks! @HaiShaw @bingxche @yctseng0211
Modifications
New test files
test/registered/amd/accuracy/mi30x/test_glm5_eval_amd.py— GSM8K few-shot completion benchmark for MI325/MI300Xtest/registered/amd/accuracy/mi35x/test_glm5_eval_mi35x.py— GSM8K few-shot completion benchmark for MI35xWorkflow updates
nightly-test-amd.yml— Addednightly-8-gpu-glm5(MI30x) andnightly-8-gpu-mi35x-glm5(MI35x) jobsnightly-test-amd-rocm720.yml— Addednightly-8-gpu-glm5-rocm720(MI30x) andnightly-8-gpu-mi35x-glm5-rocm720(MI35x) jobsConfiguration
zai-org/GLM-5(TP=8)--nsa-prefill-backend tilelang --nsa-decode-backend tilelang--chunked-prefill-size 131072,--mem-fraction-static 0.80--watchdog-timeout 1200(20 min for weight loading)Accuracy Tests
GLM-5 Models (MI325)
GLM-5 Models (MI35x)
Benchmarking and Profiling
Checklist
Review Process
/tag-run-ci-label,/rerun-failed-ci,/tag-and-rerun-ci