[Bugfix][Attention] Fix FlashInfer MLA block size logic

[MatthewBonanni]

Purpose

Before, specifying the FLASHINFER_MLA backend without a specified block size would lead to error. Block size would default to 16, and the backend only supports 32 or 64. This PR fixes it by overriding in a manner similar to the CUTLASS_MLA backend

Test Plan

VLLM_ATTENTION_BACKEND=FLASHINFER_MLA vllm bench throughput --model=deepseek-ai/DeepSeek-V2-Lite-Chat --dataset-name=random --input-len=128 --output-len=128 --num-prompts=100 --kv-cache-dtype=auto

Test Result

(no error, block size set automatically to 64)

Throughput: 28.96 requests/s, 7411.13 total tokens/s, 3706.87 output tokens/s
Total num prompt tokens:  12791
Total num output tokens:  12800

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[gemini-code-assist]

Code Review

This pull request correctly fixes a bug where using the FLASHINFER_MLA backend without specifying a block size would cause an error. The changes ensure that a supported block size (64) is automatically selected, similar to how other MLA backends are handled. The changes in check_and_update_config are correct and directly address the issue. The logic for auto-selecting the FLASHINFER_MLA backend in get_attn_backend_cls is also a good addition. I have one suggestion to improve the future-proofing for the auto-selection logic. Overall, this is a good fix.

[gemini-code-assist]

The use of cls.is_device_capability(100) for auto-selecting the FlashInfer MLA backend is too restrictive. It will only match for devices with exactly compute capability 10.0 (Blackwell), and will not automatically select this backend for future architectures with higher compute capabilities (e.g., > 10.0).

The corresponding test for this kernel (tests/kernels/attention/test_flashinfer_mla_decode.py) uses current_platform.has_device_capability(100), which suggests the kernel is expected to work on compute capabilities 10.0 and above.

To ensure future compatibility and correct auto-selection on upcoming hardware, cls.has_device_capability(100) should be used instead. This will match devices with compute capability 10.0 or greater.

A similar issue exists for the cutlass_mla backend logic, which you may want to address in a separate change for consistency.

Suggested change

	use_flashinfermla = selected_backend == _Backend.FLASHINFER_MLA or (
	selected_backend is None and cls.is_device_capability(100)
	and block_size in [32, 64])
	use_flashinfermla = selected_backend == _Backend.FLASHINFER_MLA or (
	selected_backend is None and cls.has_device_capability(100)
	and block_size in [32, 64])

[LucasWilkinson]

LGTM; thanks!