[4/N][Refactor] Refactor AscendAttentionMetadataBuilder for better extensibility and make the builder class of torchair extend from it

[shen-shanshan]

What this PR does / why we need it?

Refactor AscendAttentionMetadataBuilder for better extensibility and make the builder class of torchair extend from it.

Extract _assemble_build_info() and _assemble_attn_metadata() method from build() in AscendAttentionMetadataBuilder for better extensibility.

Workflow of build() method:

• Prepare build info: the common logic of preparing build info.
• _assemble_build_info(): the custom logic that can be overwritten in torchair_attention.py.
• _assemble_attn_metadata(): the custom logic that can be overwritten in torchair_attention.py.

After this refactor, we can remove the build() method in AscendAttentionTorchairMetadataBuilder, and just need to overwrite these two methods: _assemble_build_info() and _assemble_attn_metadata().

Note

Do not merge this PR before #2017.

Does this PR introduce any user-facing change?

How was this patch tested?

• vLLM version: v0.10.1.1
• vLLM main: vllm-project/vllm@67c1490

[gemini-code-assist]

Code Review

This pull request refactors the build method in AscendAttentionMetadataBuilder by extracting _prepare_build_info and _assemble_build_info. This is a good change that improves modularity and extensibility, as demonstrated by the new AscendAttentionTorchairMetadataBuilder. The implementation is solid, but I've found one area in the new torchair_attention.py file with some confusing code that could be clarified.

[gemini-code-assist]

The variable pad_value is assigned the value 0 on line 247, but this value is never used because it is unconditionally reassigned to 1 on line 252 before its first use. This makes the assignment on line 247 dead code, which is confusing and should be removed to improve clarity.

                    num_token_pad_size = graph_pad_size - num_actual_tokens
                    num_reqs_pad_size = (
                        graph_pad_size // self.runner.decode_token_per_req -
                        num_reqs)

[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.

[github-actions]

This pull request has conflicts, please resolve those before we can evaluate the pull request.

[codecov]

Codecov Report

❌ Patch coverage is 92.00000% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 78.01%. Comparing base (5d8ec28) to head (cafae88).

Files with missing lines	Patch %	Lines
vllm_ascend/attention/attention_v1.py	92.00%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2375      +/-   ##
==========================================
+ Coverage   77.99%   78.01%   +0.02%     
==========================================
  Files         134      134              
  Lines       18498    18515      +17     
==========================================
+ Hits        14427    14444      +17     
  Misses       4071     4071              

Flag	Coverage Δ
unittests	78.01% <92.00%> (+0.02%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.