[Test] Revert #16824 to isolate DeepSeek-V4-Pro Nightly regression

[pgzddxx]

Purpose

Experimental full revert of #16824 (merge commit e139b7d) on current upstream main d76a8c6. This PR is for isolating the DeepSeek-V4-Pro-w4a8-prefix-cache-PD Nightly performance regression. It is not a confirmed fix; please do not merge based on this PR alone.

Test request

Please run the named Nightly case on this branch and compare it with the same case on the pinned main baseline. Keep model weights, workload, hardware, image build settings, and benchmark parameters identical. Record the full AISBench table, Output TPS, Prefill TPS, TTFT, TPOT, and service logs. This PR must be evaluated independently of the #15514 revert PR; do not stack the two reverts.

Scope and risk

This reverts the arch22 QLI v2 K-scale block stride from int64_t to int32_t and removes the matching high-physical-page regression test. The original #16824 fixes signed-32-bit offset overflow for large page IDs. Reverting it can reintroduce incorrect results for that case, so the PR is strictly diagnostic and should not be merged as a production fix without addressing correctness. The target PD Nightly YAML is unchanged by this revert.

Validation

• Reverted test file: Python AST parse passed.

• Git diff whitespace check passed.

• No NPU service or benchmark has been run locally. Correctness and performance remain unverified; CI is the intended experiment.

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

[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]

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 performs an experimental revert of a previous change to the arch22 QLI v2 kernel. The goal is to isolate a performance regression observed in the DeepSeek-V4-Pro nightly benchmarks. This change is strictly for diagnostic purposes to verify if the previous update to the stride calculation is the source of the performance degradation.

Highlights

• Kernel Logic Reversion: Reverted the QLI v2 K-scale block stride calculation from int64_t back to int32_t in the arch22 kernel to isolate performance regressions.
• Test Removal: Removed the associated high-physical-page regression test in the test suite to match the reverted kernel logic.

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. ↩

[gemini-code-assist]

Code Review

Suggested PR Title:

[Attention][Misc] Revert kScaleBlkStride to int32_t and remove high physical page test

Suggested PR Summary:

### What this PR does / why we need it?
This pull request reverts `kScaleBlkStride` from `int64_t` to `int32_t` in `quant_lightning_indexer_v2_service_vector_arch22.h` and removes the regression test `test_quant_lightning_indexer_v2_high_physical_page_scale_offset` from `test_quant_lightning_indexer_v2.py`.

However, reverting `kScaleBlkStride` to `int32_t` reintroduces a critical signed 32-bit integer overflow vulnerability. When `blockId` is multiplied by `kScaleBlkStride`, the product can exceed `2^31 - 1` for large page IDs, leading to undefined behavior and incorrect memory indexing on `kScaleGm`. This change should not be merged.

### Does this PR introduce _any_ user-facing change?
No.

### How was this patch tested?
The regression test for high physical page scale offset was removed, and no new tests were added.

[gemini-code-assist]

Reverting kScaleBlkStride to int32_t reintroduces a critical signed 32-bit integer overflow vulnerability. When blockId is multiplied by kScaleBlkStride, the product can exceed 2^31 - 1 for large page IDs, leading to undefined behavior and incorrect memory indexing on kScaleGm. This change must not be merged into production.

        int64_t kScaleBlkStride = constInfo_.keyDequantScaleStride0 != 0 ? static_cast<int64_t>(constInfo_.keyDequantScaleStride0) : kCacheBlockSize_;

[wenjun91]

/nightly DeepSeek-V4-Pro-w4a8-prefix-cache-PD
nightly command triggered.

• Nightly-A3: workflow run