[Performance] Qwen3-Next: speed up update_mamba_state_after_mtp_verify by 10x; e2e up to 3.54% faster

[byjiang1996]

Modifications

Before: avg 3157.98 us

After: avg 307.05 us - 10X FASTER

Accuracy Tests - GSM8k

python3 benchmark/gsm8k/bench_sglang.py 

Accuracy remains ~0.950 before and after

Benchmark - GSM8k

Config	BEFORE (tokens/s)	AFTER (tokens/s)	% Faster
TP4 MTP	2055.889	2103.981	+2.34%
TP4 DP2 MTP	1858.580	1924.054	+3.53%

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.

[gemini-code-assist]

Summary of Changes

Hello @byjiang1996, 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 focuses on a critical performance optimization for the Qwen3-Next model by refactoring the update_mamba_state_after_mtp_verify function. The primary goal is to accelerate the state update mechanism, which has been achieved by transitioning from a loop-based, element-wise processing approach to highly efficient vectorized tensor operations. This optimization not only drastically speeds up the specific function but also contributes to a measurable improvement in the model's overall inference speed, all while ensuring that the model's accuracy remains unaffected.

Highlights

• Performance Improvement: The update_mamba_state_after_mtp_verify function has been optimized to run 10 times faster, reducing its average execution time from 3157.98 us to 307.05 us.
• End-to-End Speedup: The overall end-to-end performance of the Qwen3-Next model has improved by up to 3.54% (specifically 3.53% for TP4 DP2 MTP configuration) on the GSM8k benchmark.
• Accuracy Maintained: Accuracy tests on GSM8k confirm that the model's performance remains consistent at approximately 0.950, indicating no degradation due to the speed optimizations.
• Vectorized State Updates: The core change involves replacing iterative, chunk-based updates for ssm_states and conv_states with direct, vectorized tensor assignments, significantly reducing Python overhead.

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 significantly improves the performance of update_mamba_state_after_mtp_verify by replacing iterative, element-wise state updates with vectorized PyTorch operations. This is a great optimization that leads to a more concise and much faster implementation. My only concern is the potential increase in peak memory usage, as the previous chunking mechanism, which was likely in place to manage memory, has been removed. I've added a comment with a suggestion to reintroduce chunking in a vectorized way, which could provide a good balance between performance and memory consumption if needed.