Fix the bug that total expert num greater than 256 or local expert num is less than 8

[luanyundu]

Tested by local expert num is 1.
Tested by total expert num is 512 and topk is 12.

Co-worker

• @luanyundu Debugging work
• @oagniqgnat Debugging work
• @WSEmma Main code modification

[gemini-code-assist]

Summary of Changes

Hello @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 and resolves bugs related to the Mixture-of-Experts (MoE) system, specifically when the total number of experts exceeds 256 or when the local expert count is less than 8. The changes involve refining the batching strategy and enhancing memory alignment and indexing logic to ensure robust and correct operation under these specific conditions, thereby improving the stability and performance of the MoE implementation.

Highlights

• Batching Logic Adjustment: Introduced a new constant EXPERT_NORMAL_NUM (256) and modified the BATCH_ROUND constant from 32 to 16. The batchRounds variable is now dynamically set to 16 if the total number of experts exceeds EXPERT_NORMAL_NUM, otherwise it defaults to 32, addressing issues with large expert counts.
• Memory Alignment and Indexing Refinement: Refactored memory alignment and indexing calculations across several methods within the NotifyDispatch class. New variables like alignedDataLen and strideElem were introduced to ensure proper memory handling and prevent issues related to data access, particularly when dealing with varying local expert numbers.
• Removed Redundant Functionality: The ReorderSendTokensPerRankOutput function was removed from the NotifyDispatch class, streamlining the codebase.
• Log Message Correction: Corrected a log message in cam_moe_combine_normal_tiling.cc from indicating 'greater than' to 'not equal to' for a dimension check, improving diagnostic clarity.

🧠 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

• csrc/deepep/ops/op_host/cam_moe_combine_normal_tiling.cc
  • Corrected the error message in CheckTensorShape from 'x's dim0 is greater than bs' to 'x's dim0 not equal to bs'.
• csrc/deepep/ops/op_kernel/notify_dispatch.h
  • Added EXPERT_NORMAL_NUM constant with a value of 256.
  • Changed BATCH_ROUND constant from 32 to 16.
  • Modified the initialization of batchRounds to be conditional based on numExperts exceeding EXPERT_NORMAL_NUM.
  • Introduced alignedDataLen and strideElem variables in DataCopyPad and related loops to correctly calculate destination offsets and pair_idx for aligned memory access.
  • Removed the ReorderSendTokensPerRankOutput method.
  • Updated recvDataAlignLen calculation in BuildTotalRecvTokens, InitRecvCountBuf, InitRecvOffsetBuf, InitMaxBsBuf, InitRecvTokenPerExpBuf, InitSrcRankExpOffsetBuf, and InitRInSrcRankOffsetBuf to use singleRankAlignLen for improved memory alignment.

Activity

• luanyundu: Debugging work
• oagniqgnat: Debugging work
• WSEmma: Main code modification

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]

Code Review

This pull request addresses a bug concerning memory layout and alignment when the number of experts changes, specifically for cases with more than 256 total experts or a small number of local experts. The main fix is in csrc/deepep/ops/op_kernel/notify_dispatch.h, where data access patterns and buffer size calculations are updated to correctly handle memory alignment for data from different ranks. This is a crucial correctness fix. However, the implementation introduces significant code duplication for these new calculations. My review includes suggestions to refactor the duplicated logic to improve code maintainability.