fix grouped_topk tl.sort when numel=1

[SangChengC]

No description provided.

[gemini-code-assist]

Summary of Changes

Hello @SangChengC, 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 a targeted fix to the grouped_topk_kernel function in the fused Mixture of Experts (MoE) module. The primary purpose is to resolve an issue related to the tl.sort function when processing a single element or expert group. By making the sorting operation conditional on the number of expert groups being greater than one, the change enhances the robustness and efficiency of the kernel, preventing potential errors or redundant computations in edge cases.

Highlights

• Conditional Sorting Logic: The tl.sort operation within the grouped_topk_kernel function is now conditionally applied only when EXPERT_GROUP_NUM is greater than 1. This prevents unnecessary or potentially problematic sorting of single-element arrays.
• Bug Fix for numel=1: This change specifically addresses a reported issue where tl.sort might behave unexpectedly or be inefficient when the number of elements (numel) is 1, as indicated by the pull request title.
• Code Simplification: The intermediate sorted_group_value variable has been removed, and group_value is now directly used after the conditional sorting, streamlining the code.

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 aims to fix an issue with tl.sort when the input has only one element. However, the current implementation introduces a critical bug by overwriting the group_value variable, which corrupts the subsequent masking logic. I've provided a suggestion to fix this by preserving the original group_value while still addressing the tl.sort issue.

[gemini-code-assist]

While this change correctly avoids calling tl.sort on a single-element tensor, it introduces a bug by overwriting group_value. The original group_value is needed on line 166 for calculating mask_group_scores. By reassigning group_value to its sorted version, the masking logic becomes incorrect.

To fix this, you should use a separate variable for the sorted values, similar to the original code, while still conditionally calling tl.sort.

    sorted_group_value = group_value
    if EXPERT_GROUP_NUM > 1:
        sorted_group_value = tl.sort(group_value, descending=True)
    group_topk_value = tl.sum(tl.where(offs_group == group_topk_num - 1, sorted_group_value, 0.0))