Revert "Fix spec decoding acc length for dpsk-r1-fp4 tp8"

[Qiaolin-Yu]

Reverts #12896

[gemini-code-assist]

Summary of Changes

Hello @Qiaolin-Yu, 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 performs a revert of a previous change that adjusted the max_token_num parameter within the flashinfer_comm_fusion module. The original modification was intended to address a specific decoding accuracy length issue, but this revert restores the max_token_num to its higher default, effectively undoing that particular adjustment and returning the system to its prior configuration.

Highlights

• Revert Previous Change: This pull request reverts the changes introduced in sgl-project/sglang#12896, which aimed to fix specific decoding issues related to dpsk-r1-fp4 tp8.
• Max Token Number Adjustment: The max_token_num parameter in both ensure_workspace_initialized and flashinfer_allreduce_residual_rmsnorm functions has been reverted from 2048 back to its previous default value of 16384.

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 reverts the default value for max_token_num in python/sglang/srt/layers/flashinfer_comm_fusion.py. This change appears to be a straightforward revert, likely to address an issue with the previous value. I have one suggestion regarding code maintainability by replacing the repeated magic number with a constant.

[gemini-code-assist]

The default value for max_token_num (16384) is used in this function, flashinfer_allreduce_residual_rmsnorm, and fake_flashinfer_allreduce_residual_rmsnorm. To improve maintainability and avoid magic numbers, consider defining this value as a module-level constant (e.g., _DEFAULT_MAX_TOKEN_NUM = 16384) and using this constant in all three function definitions.

[zhyncs]

Let's add the 4 gpu and 8 gpu uts @Kangyan-Zhou @Qiaolin-Yu

[Kangyan-Zhou]

Let's add the 4 gpu and 8 gpu uts @Kangyan-Zhou @Qiaolin-Yu

@Qiaolin-Yu Shall we add 8 gpu tests to nightly tests to test/srt/nightly and https://github.com/sgl-project/sglang/blob/main/.github/workflows/nightly-test-b200.yml?

[Qiaolin-Yu]

Let's add the 4 gpu and 8 gpu uts @Kangyan-Zhou @Qiaolin-Yu

@Qiaolin-Yu Shall we add 8 gpu tests to nightly tests to test/srt/nightly and https://github.com/sgl-project/sglang/blob/main/.github/workflows/nightly-test-b200.yml?

Sure. We should at least run a case like dpsk-r1-fp4 tp8 + overlap spec decoding