Use Blackwell FlashInfer MXFP4 MoE by default if available

[mgoin]

Purpose

Essentially we should act like VLLM_USE_FLASHINFER_MOE_MXFP4_BF16=1 if on SM100 and flashinfer is installed

Test Plan

Test Result

Main branch:

lm_eval --model vllm --model_args pretrained=openai/gpt-oss-20b --trust_remote_code --tasks gsm8k --num_fewshot 5 --batch_size auto
(EngineCore_0 pid=4179585)   File "/home/mgoin/code/vllm/vllm/model_executor/layers/quantization/mxfp4.py", line 356, in process_weights_after_loading
(EngineCore_0 pid=4179585)     from triton_kernels.matmul_ogs import FlexCtx, PrecisionConfig
(EngineCore_0 pid=4179585) ModuleNotFoundError: No module named 'triton_kernels'

VLLM_USE_FLASHINFER_MOE_MXFP4_BF16=1 lm_eval --model vllm --model_args pretrained=openai/gpt-oss-20b --trust_remote_code --tasks gsm8k --num_fewshot 5 --batch_size auto
vllm (pretrained=openai/gpt-oss-20b,trust_remote_code=True), gen_kwargs: (None), limit: None, num_fewshot: 5, batch_size: auto
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.3647|±  |0.0133|
|     |       |strict-match    |     5|exact_match|↑  |0.2328|±  |0.0116|

This PR:

lm_eval --model vllm --model_args pretrained=openai/gpt-oss-20b --trust_remote_code --tasks gsm8k --num_fewshot 5 --batch_size auto
vllm (pretrained=openai/gpt-oss-20b,trust_remote_code=True), gen_kwargs: (None), limit: None, num_fewshot: 5, batch_size: auto
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.3647|±  |0.0133|
|     |       |strict-match    |     5|exact_match|↑  |0.2328|±  |0.0116|

(Optional) Documentation Update

---

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.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[gemini-code-assist]

Code Review

This pull request enables the Blackwell FlashInfer MXFP4 MoE backend by default when available. The changes introduce helper functions to centralize the logic for determining which FlashInfer backend to use and refactor the existing code to call these helpers. A helpful warning is also added for users on Blackwell hardware when FlashInfer is not installed. The overall changes are well-structured. I've identified one high-severity issue regarding the prioritization of FlashInfer backends, which could lead to using a less performant kernel even when a faster one is enabled. A code suggestion is provided to address this.

[yewentao256]

LGTM, thanks for the work!

[celsowm]

sm120 / rtx 50xx gonna finally works too ?

[SabareeshGC]

There are sm120 such as RTX 6000 Pro Blackwell , shouldn't we include those as well ?