[Core] Disable the chunked prefill feature in Non-MLA LLMs

[rjg-lyh]

What this PR does / why we need it?

This PR enforces the forcible disabling of the chunked prefill feature in Non-MLA models, as the performance of operators supporting this functionality is currently suboptimal. Unless the user has enabled chunked prefill in the ascend_scheduler_config, we would allow this feature.

Does this PR introduce any user-facing change?

No.

How was this patch tested?

CI passed with new added/existing test.

Related: #2659

• vLLM version: main
• vLLM main: vllm-project/vllm@d21a36f

[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

This pull request aims to disable chunked prefill for Non-MLA models by enabling the AscendScheduler, which has this feature off by default. The overall logic is sound. I've provided two high-severity comments. The first addresses a misleading log message that could cause confusion. The second suggests a refactoring to simplify a verbose and potentially brittle boolean check, improving code clarity and robustness.

[gemini-code-assist]

The log message is misleading. It states that chunked prefill is "forcibly disabled", but the code only enables a scheduler that has it disabled by default and warns the user if they have explicitly enabled it. This can be confusing for users. I suggest a more accurate and concise message.

Suggested change

	"Non-MLA LLMs forcibly disable the chunked prefill feature,"
	"as the performance of operators supporting this feature "
	"functionality is currently suboptimal.")
	"For Non-MLA models, chunked prefill is disabled by default for performance reasons."

[gemini-code-assist]

The logic to check if chunked prefill is enabled is verbose and uses a brittle == True comparison, which can lead to unexpected behavior if the configuration value is not a strict boolean. This can be simplified and made more robust by using getattr and a direct boolean check, which improves readability and correctness.

                chunked_prefill_enabled_in_ascend_scheduler = getattr(
                    ascend_scheduler_config, "enable_chunked_prefill", False)
                if chunked_prefill_enabled_in_ascend_scheduler:
                    logger.warning(
                        "Chunked prefill feature is enabled in ascend_scheduler,"
                        "but note that the operator supporting this feature "
                        "would lead to performance degradation.")

[codecov]

Codecov Report

❌ Patch coverage is 83.33333% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 74.97%. Comparing base (1bbb20e) to head (65013d9).
⚠️ Report is 31 commits behind head on main.

Files with missing lines	Patch %	Lines
vllm_ascend/platform.py	83.33%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2894      +/-   ##
==========================================
+ Coverage   74.76%   74.97%   +0.20%     
==========================================
  Files         150      154       +4     
  Lines       20891    21308     +417     
==========================================
+ Hits        15620    15976     +356     
- Misses       5271     5332      +61     

Flag	Coverage Δ
unittests	74.97% <83.33%> (+0.20%)	⬆️

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.

[Yikun]

The round2 perf has perf regression which is very wired, but compare to v0.10.1rc1 it's OK, so merge this first.