[FlashInfer] Avoid FlashInfer block_size 16 + head_size 256 on blackwell

[heheda12345]

Purpose

Avoid this combination as flashinfer-ai/flashinfer#1993 reports this combination is not correct.

For most models with head_size 256, users now need --block_size 32 / --block_size 64

For hybrid mamba like qwen3-next, the block_size can be resolved automatically

Thanks @vadiklyutiy for the exploration on this problem #27704

Test Plan

Test non-hybrid case and hybrid case

Test Result

All tests are on B200

Normal model

I don't know a model with head_size 256. I've changed the 256 in this PR to 64 and test opt-125m

vllm serve facebook/opt-125m

will result in

AssertionError: There is a bug in FlashInfer block_size 16 head size 256 support. Please avoid this combination by passing --block-size 32 or --block-size 64.

vllm serve facebook/opt-125m --block-size 64

works

Hybrid

vllm serve Qwen/Qwen3-Next-80B-A3B-Instruct -tp 4
lm_eval --model local-completions --model_args model=Qwen/Qwen3-Next-80B-A3B-Instruct,base_url=http://localhost:8000/v1/completions -t gsm8k --num_fewshot 5 --batch_size 250

before this PR

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.2123|±  |0.0113|
|     |       |strict-match    |     5|exact_match|↑  |0.1933|±  |0.0109|

after this PR

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.8491|±  |0.0099|
|     |       |strict-match    |     5|exact_match|↑  |0.8120|±  |0.0108|

---

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.

[MengqingCao]

nit: How about make the logic of kernel_block_alignment_size into each attention backends? Actually this is need in vllm-ascend, too. If this is acceptable, I could help to make a follow-up pr

[heheda12345]

Yes it make sense. Help wanted on it. And I think the alignment can be inferred from get_supported_kernel_block_size

[MengqingCao]

Okay, I'll work on it then.

[pavanimajety]

For my knowledge, what is kernel_block_alignment_size?

[vadiklyutiy]

@MengqingCao
are you talking about something like I did in #27704

[MengqingCao]

@MengqingCao are you talking about something like I did in #27704

Yes, thanks for the context! I'll take a look at your pr

[pavanimajety]

LGTM, we can undo the change once flashinfer is fixed.