GPT-OSS Support: Add Blackwell MoE mxfp4 implementation from TRTLLM and Attention Sink #1389
Conversation
joker-eph
force-pushed
the
moe_and_attention_sink
branch
from
b044dff to
809b8d7
Compare
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Summary of Changes
Hello @joker-eph, 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 introduces significant advancements in low-precision computation and attention mechanisms. I've implemented the Blackwell MXFP8 quantization, providing new kernels and PyTorch interfaces for efficient mixed-precision operations. Concurrently, I've integrated 'Attention Sinks' into the paged attention kernels, which will enhance the model's ability to manage long sequence contexts. Furthermore, I've undertaken a substantial refactoring of the Mixture of Experts (MoE) system, separating routing logic into distinct, optimized kernels and updating the underlying GEMM and MoE runners to support a broader range of data types and activation controls. These changes collectively aim to improve performance, flexibility, and architectural alignment.
Highlights
- Blackwell MXFP8 Quantization: I've integrated the Blackwell MXFP8 quantization implementation, including new CUDA kernels and PyTorch bindings, to enable efficient low-precision operations.
- Attention Sink Implementation: I've added support for 'Attention Sinks' in both decode and context attention kernels, allowing for improved handling of long sequences by modifying the softmax denominator.
- MoE Routing Kernel Refactoring: I've significantly refactored the Mixture of Experts (MoE) routing kernels, separating them into specialized DeepSeek, Llama4, and Renormalize implementations for better modularity and performance.
- Enhanced Batched GEMM and MoE Runner: I've updated the batched GEMM runner and MoE runner to support new data types, activation types, and parameters for bias, alpha, beta, and clamp limits, enhancing flexibility and control over fused operations.
- Blackwell MoE Performance Optimizations: I've introduced vectorized loading optimizations in the MoE finalize kernel, specifically targeting the Blackwell architecture for improved performance.
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 in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.
| 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 issue 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 or fill out our survey to provide feedback.
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
-
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. ↩
There was a problem hiding this comment.
Code Review
This pull request introduces significant new features and refactorings, primarily adding support for Blackwell architecture's MoE mxfp4 implementation and Attention Sink. The changes are extensive, touching quantization kernels, fused MoE components, and attention launchers. The MoE and GEMM runners have been refactored to be more generic, supporting a wider range of data types and fused operations. While the changes are generally well-structured, I've identified a couple of areas where type safety and code clarity could be improved.
|
@yzh119 : I pushed an extra commit to address Gemini's comment, but I haven't tested it yet. Let me know if I should pull it out of this PR. |
|
This macro seems in conflict with |
IwakuraRein
force-pushed
the
moe_and_attention_sink
branch
from
d1d1918 to
5587906
Compare
joker-eph
changed the title
<!-- .github/pull_request_template.md --> ## 📌 Description Bump version to v0.2.10 for GPT-OSS Support: #1389 ## 🔍 Related Issues - #1389 - https://openai.com/index/introducing-gpt-oss/ ## 🚀 Pull Request Checklist Thank you for contributing to FlashInfer! Before we review your pull request, please make sure the following items are complete. ### ✅ Pre-commit Checks - [x] I have installed `pre-commit` by running `pip install pre-commit` (or used your preferred method). - [x] I have installed the hooks with `pre-commit install`. - [x] I have run the hooks manually with `pre-commit run --all-files` and fixed any reported issues. > If you are unsure about how to set up `pre-commit`, see [the pre-commit documentation](https://pre-commit.com/). ## 🧪 Tests - [x] Tests have been added or updated as needed. - [x] All tests are passing (`unittest`, etc.). --------- Co-authored-by: yzh119 <[email protected]>
These kernels support OpenAI GPT-OSS
Co-authored-by: siyuanf [email protected]
Co-authored-by: Zihao Ye [email protected]
Co-authored-by: Qidi Sang [email protected]