Conversation
Summary of ChangesHello @luanyundu, 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 a hanging bug and enhances the performance of MoE operations, especially when dealing with larger batch sizes. The core changes involve increasing the maximum allowed batch size and refining the multi-round processing logic to ensure proper synchronization and data handling, which is crucial for stable and efficient execution. 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. 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
|
There was a problem hiding this comment.
Code Review
This pull request effectively addresses a hanging bug and improves performance, as indicated by the provided benchmark data. The changes involve increasing the BS_UPPER_BOUND constant in two tiling configuration files to support larger batch sizes and a significant reordering of synchronization logic within the CamMoeCombineNormalMultiRound::Process method. The reordering of operations in the Process method appears to be a well-considered fix, ensuring buffer operations are performed consistently before conditional inter-round synchronization.
…-npu into sgl-cmake2 * 'sgl-cmake2' of https://github.com/1329009851/sgl-kernel-npu: CI execution requirements for separating a2 and a3 (sgl-project#367) Fix the bug that total expert num greater than 256 or local expert num is less than 8 (sgl-project#364) adapt ant moving to A2 single machine (sgl-project#362) reset ci -- run test mixed running for experts on a2. (sgl-project#365) Revert "Build the deepep package with the chip model included. (sgl-project#274)" (sgl-project#363) fix:buffer control (sgl-project#361) Build the deepep package with the chip model included. (sgl-project#274) bugfix wrong packages build dir (sgl-project#360) bump version to 2026.02.01 (sgl-project#359) Cover the workflows cases on a3 (sgl-project#321) release follows naming convention (sgl-project#356) Modify notifydispatch to support DEEPEP_NORMAL_LONG_SEQ_ROUND up to 128. (sgl-project#352) fix the hanging bug (sgl-project#355) [Bugfix] Fix build script working with cann 8.5.0 (sgl-project#354) Modify the description of DeepEP in the README file. (sgl-project#348) Revert "Add scripts for building CMake files (sgl-project#344)" (sgl-project#353) Add scripts for building CMake files (sgl-project#344) Support x86_64 and aarch64 binary release (sgl-project#325) add function for deep-ep tests (sgl-project#301) [Doc] Improved README.md content and English grammar and integrated the DeepWiki badge for Ask AI (sgl-project#345)
the performance is below