Replace clamp_position with JIT kernel + platform dispatch#20999
Replace clamp_position with JIT kernel + platform dispatch#20999merrymercy merged 5 commits intomainfrom
Conversation
|
/tag-and-rerun-ci |
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 significantly enhances the performance of the 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. 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. Footnotes
|
There was a problem hiding this comment.
Code Review
This pull request introduces a custom JIT CUDA kernel for clamp_position to improve performance, along with a platform dispatch mechanism. The changes are well-structured, including the kernel, Python wrappers, comprehensive unit tests, and benchmarks demonstrating significant performance gains. My main feedback addresses a data type inconsistency between the new CUDA implementation and the implementations for other platforms, which could lead to subtle bugs. I've suggested changes to ensure the CUDA path consistently returns int64 tensors, aligning it with other platforms and the original behavior.
- Add a CUDA JIT kernel for clamp_position (clamp(seq_lens - 1, min=0)) - Use JIT kernel on CUDA, torch.compile on HIP/AMD, plain torch elsewhere - Add unit tests (64 cases) and benchmark Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Documents the full workflow: launch server, validate accuracy, capture Chrome-compatible trace, kill server, report profile path. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
dca3ffc to
666b601
Compare
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…ct#20999) Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…ct#20999) Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…ct#20999) Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…ct#20999) Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Summary
clamp_position.cuh) for computing decode positions:clamp(seq_lens - 1, min=0), templated on dtype (int32/int64)@cache_oncefor module caching@torch.compileimplementation inforward_batch_info.pywith platform dispatch: CUDA JIT kernel on CUDA,torch.compileon HIP/AMD, plain torch on other platforms (NPU, CPU, etc.)Benchmark (NVIDIA H100)
Test plan
python -m pytest python/sglang/jit_kernel/tests/test_clamp_position.py -v -s(64/64 passed)python -m pytest test/registered/core/test_srt_engine.py -x -v(8/8 passed)🤖 Generated with Claude Code