[Bugfix][Revert][Attention] Revert AscendC MsaIndexScore MiniMax-M3 path

[yingyingzizi]

What this PR does / why we need it?

Revert PR #16426, [Performance] Enable AscendC index score decode for MiniMax-M3.

The A5 nightly validation cannot currently register or execute npu_msa_index_score. The latest A5 run reached pytest and all 24 precision cases failed because torch.ops._C_ascend.npu_msa_index_score was unavailable. Reverting the change removes the failing AscendC MiniMax-M3 index-score path until the custom-operator build and registration issue is resolved.

This PR reverts merge commit fefa33767507514577b80affaa90b5b77ce8ac8a and preserves unrelated changes already present on main, including subsequent nightly matrix maintenance.

Does this PR introduce any user-facing change?

Yes. MiniMax-M3 index-score execution returns to the pre-#16426 implementation. The AscendC MsaIndexScore optimization and its associated source, documentation, integration changes, and precision test are removed.

How was this patch tested?

• Reverted merge commit fefa33767507514577b80affaa90b5b77ce8ac8a with parent -m 1.
• Resolved conflicts by preserving current main workflow changes and removing only the [Performance] Enable AscendC index score decode for MiniMax-M3 #16426 changes.
• git diff --check passed.
• The revert commit includes a Signed-off-by line.
• A5 nightly failure: https://github.com/vllm-project/vllm-ascend/actions/runs/35500139842/job/106055635575

Related: #16426, #16955, #17013.

• vLLM main: vllm-project/vllm@84030bb

[gemini-code-assist]

Summary of Changes

Hello, 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 reverts the recent addition of the AscendC-based index score decoding path for MiniMax-M3. The revert was necessary because the custom operator failed to register correctly in the A5 nightly validation environment, causing all precision test cases to fail. The system now falls back to the previous implementation, ensuring stability while the underlying build and registration issues are addressed.

Highlights

• Revert AscendC MiniMax-M3 Path: Reverted the integration of the AscendC npu_msa_index_score operator for MiniMax-M3 due to unresolved build and registration issues in the A5 nightly environment.
• Feature Removal: Removed the AscendC MsaIndexScore optimization, associated documentation, integration logic, and precision tests, reverting to the pre-PR [Performance] Enable AscendC index score decode for MiniMax-M3 #16426 implementation.
• Build and Test Cleanup: Cleaned up build scripts, test references, and model-specific logic that relied on the now-reverted AscendC operator.

New Features

🧠 You can now enable Memory (public preview) 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.

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 the 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 counterproductive. 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.

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. ↩

[github-actions]

👋 Hi! Thank you for contributing to the vLLM Ascend project. The following points will speed up your PR merge:‌‌

• A PR should do only one thing, smaller PRs enable faster reviews.
• Every PR should include unit tests and end-to-end tests ‌to ensure it works and is not broken by other future PRs.
• Write the commit message by fulfilling the PR description to help reviewer and future developers understand.

If CI fails, you can run linting and testing checks locally according Contributing and Testing.

[gemini-code-assist]

Code Review

Suggested PR Title:\n\nmarkdown\n[Attention][Misc] Simplify MsaIndexScore tiling and remove replicated decode logic\n\n\nSuggested PR Summary:\n\nmarkdown\n### What this PR does / why we need it?\nThis PR simplifies the `MsaIndexScore` operator by removing the complex `kvChunks` and `launchAic` estimation logic from the host tiling, and removing the replicated decode logic (`minimax_m3_index_decode_replicated`) for A5. It cleans up the `MsaTask` structure and simplifies the epilogue logic for both arch22 and arch35.\n\nHowever, a critical bug was introduced in the arch22 epilogue (`msa_index_score_epilogue.h`) where the removal of `OpenStageWindow` and `AdvanceStageWindow` leaves `stageBlkEnd_` uninitialized, causing `FlushStage()` to always return early and break attention score outputs on Atlas A2/A3 during decode.\n\n### Does this PR introduce _any_ user-facing change?\nNo user-facing API changes are introduced, but it simplifies internal scheduling and removes unused replicated decode paths.\n\n### How was this patch tested?\nThe PR adds a comprehensive end-to-end accuracy self-check test suite in `test_aclnn_msa_index_score.cpp` covering 40 test cases on Ascend 950 and 36 cases on A2/A3.\n

[yingyingzizi]

/rerun
[Bot]: rerun completed.

Rerun (failed jobs only):

• E2E