Update Flashinfer to 0.2.14.post1

[weireweire]

Purpose

Update flashinfer to

• fix allreduce fusion kernel suboptimal perf issue.
• Add GPT-OSS cutlass MoE backend.
• Include [Do Not Merge]Fix after flashinfer fp4 autotuner pr #23209 for flashinfer autotune @IwakuraRein
• Include Draft: Fix flashinfer swizzle enum name for flashinfer update. #23311 for flashinfer enum API change.

Test Plan

lm_eval on llama3

gpt-oss/eval test on gpt-oss

• python3 -m gpt_oss.evals --sampler chat_completions --model gpt-oss-120b --reasoning-effort low,medium --n-threads 512 --eval gpqa

Test Result

llama3 FP8 tp2

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
gsm8k	3	flexible-extract	5	exact_match	↑	0.938	±	0.0108
		strict-match	5	exact_match	↑	0.904	±	0.0132

llama3 FP4 tp1

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
gsm8k	3	flexible-extract	5	exact_match	↑	0.930	±	0.0114
		strict-match	5	exact_match	↑	0.842	±	0.0163

gpt-oss tp1:
[{'eval_name': 'gpqa', 'model_name': 'gpt-oss-120b-low_temp1.0_20250825_021150', 'metric': 0.6414141414141414}, {'eval_name': 'gpqa', 'model_name': 'gpt-oss-120b-medium_temp1.0_20250825_021150', 'metric': 0.711489898989899}]

(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.

[gemini-code-assist]

Code Review

This pull request updates Flashinfer to version 0.2.14.post1, which addresses a performance issue in the allreduce fusion kernel and incorporates API changes. The changes also include enabling FlashInfer autotuning before CUDA graph capture for better performance. My review found a critical typo in a variable name (max_captute_size instead of max_capture_size) in mxfp4.py which would lead to a runtime error. I've provided suggestions to fix this.

[yewentao256]

Thanks for the work!
Could you also report vllm bench metric results so that we can see if we have some improvement for E2E throughput?

[mgoin]

LGTM, thanks for putting everything together. Let's see the CI