[Feature][Model] Integrate Triton KeyPool indexing for GLM-5.3-Flash

[lijiahang226]

What this PR does / why we need it?

Refs #15665

Connect the GLM-5.3-Flash pooled indexer to IndexerWrapper and the shared NoPE SFA backend. Keep cache metadata and the model-side indexer backend together in vllm_ascend/attention/indexer_kpool.py. Load KPool execution dependencies when the backend is instantiated to avoid circular imports during metadata-only startup.

Keep hardware and PCP/DCP capability checks in backend initialization. Pass complete compressed physical pages and compressor-state metadata to the Triton operators while preserving causal-tail visibility and the cache allocation/grouping contract from #15913.

Dependencies:

• [Feature][Ops] Add Triton KeyPool compression and pooled indexing #16243 provides the Triton operators and indexer backend-selection hook.
• [Feature][Attention] Support NoPE in the shared SFA backend #16252 provides shared NoPE SFA execution and metadata.

Prerequisite commits are combined only for integration validation and are excluded from this branch.

Does this PR introduce any user-facing change?

Yes. GLM-5.3-Flash routes pooled sparse attention through the shared SFA backend and Triton indexer. Unsupported hardware and PCP/DCP configurations are rejected when the backend is initialized.

How was this patch tested?

At PR revision 1ff19ad48, combined with its integration prerequisites:

• Fresh-process imports of the consolidated metadata module and model backend factory passed. Regression tests cover importing metadata without generic SFA indexing or KPool execution dependencies.
• 120 unit tests passed, 1 skipped, covering test_glm5next_kpool_model_backend.py, test_glm5next_kv_cache.py, test_glm5next_sfa_routing.py, test_glm5next_indexer_backend.py, test_platform.py, and attention/test_indexer.py.
• Six native NPU comparisons against the pre-consolidation backend produced identical indices and compressed/state caches: prefill tail, decode tail, pool-boundary decode, prefill top-k selection, FULL-mode padding, and cache-only updates.
• GitHub CI passed pre-commit, CPU unit tests, selected device tests, and the CI gate.

Earlier full-model validation at PR revision 0bff0eb46, combined with integration prerequisites, passed 320 unit tests (1 skipped) and 8/8 GLM-5.3-Flash-w8a8 inference cases with TP8/EP8, FULL_DECODE_ONLY, and MTP=3. Those results apply to the pre-consolidation revision; the six native comparisons above do not constitute a new full-model or ACLGraph capture/replay run.

• vLLM main: vllm-project/vllm@b2f6858

[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 integrates the GLM-5.3-Flash pooled sparse attention mechanism into the vLLM Ascend backend. By leveraging Triton operators for KeyPool compression and routing through the shared SFA infrastructure, the changes enable efficient sparse attention on supported hardware (A2/A3/A5). The implementation simplifies cache management by utilizing complete physical pages and improves routing logic to ensure compatibility with existing serving constraints.

Highlights

• GLM-5.3-Flash KPool Integration: Integrated the GLM-5.3-Flash pooled indexer with the IndexerWrapper and shared NoPE SFA backend, enabling Triton-based persistent KeyPool compression and pooled-key selection.
• Indexer Backend Refactoring: Moved model-specific projections and cache binding into the indexer backend, removing the CANN pool-indexer block-splitting path in favor of complete physical page addressing.
• Routing and Validation: Added explicit validation to reject unsupported PCP/DCP configurations and implemented new regression tests for indexer page addressing and causal tail handling.

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.

---

Tip

💡 Consider Linking a Related Issue or RFC

Your PR title contains the [Feature] tag, indicating a bug fix or new feature.

Linking a related issue or RFC in the PR description is strongly encouraged — it gives reviewers helpful context and speeds up the review. You can use any of these keywords:

• Fixes #<issue_number>
• Closes #<issue_number>
• Resolves #<issue_number>
• Refs #<rfc_or_issue_number> (for RFCs)

🙏 Thanks for helping us keep the project well-organized!

[gemini-code-assist]

Code Review

Suggested PR Title:

[Attention][Feature] Implement GLM-Next KPool sparse indexer backend and Triton orchestration

Suggested PR Summary:

### What this PR does / why we need it?

This PR implements the model-side backend and Triton orchestration for the GLM-Next pooled sparse indexer (KPool) on Ascend NPU. It replaces the previous placeholder implementation that raised `NotImplementedError` with a fully functional Triton-based indexer.

Key changes include:
- Implementing `Glm5NextKPoolIndexerBackend` to handle the unified indexer API.
- Rewriting `SparseAttnIndexerKpool` to orchestrate Triton kernels (`glm5_next_kpool_state_compress_and_write_cache_triton` and `glm5_next_lightning_indexer_triton`).
- Adding `append_causal_tail` to handle unpooled tokens for CANN SFA compatibility.
- Updating platform routing to support KPool sparse attention on Ascend A2, A3, or A5, while explicitly rejecting context parallelism.
- Removing obsolete helper functions like `select_indexer_block_size` and the unused `fwht128_quant_fp8` module.
- Adding comprehensive unit and end-to-end tests to verify indexer pages, causal tail, and backend routing.

### Does this PR introduce _any_ user-facing change?

No, this is an internal backend implementation for GLM-Next sparse attention indexing on Ascend NPUs.

### How was this patch tested?

Tested with new unit and end-to-end tests:
- `test_glm5next_indexer_pages.py`
- `test_glm5next_kpool_tail.py`
- `test_glm5next_kpool_model_backend.py`
- `test_glm5next_sfa_routing.py`

***

*Note: No review comments were provided for this pull request, so there is no additional feedback to address.*

[pisceskkk]

I noticed that the vllm_ascend/attention directory is currently taking on too many responsibilities. The newly added indexer_kpool_backend.py does not seem substantial enough to warrant a separate file, so we can temporarily merge it into indexer_kpool.py.

[lijiahang226]

I noticed that the vllm_ascend/attention directory is currently taking on too many responsibilities. The newly added indexer_kpool_backend.py does not seem substantial enough to warrant a separate file, so we can temporarily merge it into indexer_kpool.py.

Sounds good. These two files have been merged.