[Kernel] Fix fused_gdn_gating

[ZJY0516]

Purpose

Fix fused_gdn_gating accuracy issue

Test Plan

vllm serve Qwen/Qwen3-Next-80B-A3B-Instruct --enable-expert-parallel -tp 4

lm_eval --model local-chat-completions --model_args model=Qwen/Qwen3-Next-80B-A3B-Instruct,base_url=http://localhost:8000/v1/chat/completions,num_concurrent=280 --tasks gsm8k --apply_chat_template --num_fewshot 5

Test Result

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
gsm8k	3	flexible-extract	5	exact_match	↑	0.7839	±	0.0113
		strict-match	5	exact_match	↑	0.6566	±	0.0131

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[gemini-code-assist]

Code Review

This pull request addresses an accuracy issue in the fused_gdn_gating_kernel. The changes are twofold: first, the manual sigmoid implementation is replaced with tl.sigmoid, which should provide better numerical stability. Second, the data type for storing the beta_output is corrected to match the input tensor b's data type. This change ensures that the kernel's behavior correctly mimics b.sigmoid(), resolving a precision mismatch that was likely the cause of the accuracy problem. The fix appears correct and well-targeted. I have no further suggestions.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[vadiklyutiy]

Could you add accuracy test before this PR

[ZJY0516]

Could you add accuracy test before this PR

Before this PR, sometimes gsm8k score was zero

[ZJY0516]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. What shall we delve into next?

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[vadiklyutiy]

Could you add accuracy test before this PR

Before this PR, sometimes gsm8k score was zero

what sometimes exactly mean?

• use the same cmd and result was randomly 0 or around .80
• or different cmd give different results
• ?

[ZJY0516]

Could you add accuracy test before this PR

Before this PR, sometimes gsm8k score was zero

what sometimes exactly mean?

• use the same cmd and result was randomly 0 or around .80
• or different cmd give different results
• ?

same cmd, randomly 0 or around .80