[Bugfix] fix qwen3-omni performance regression

[R2-Y]

To avoid performance degradation when using FlashInfer CUTLASS Unquantized MoE backend, Qwen3-Omni sets VLLM_USE_FLASHINFER_MOE_FP16=0 by default (Using TRITON Unquantized MoE backend).

PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS (AT THE BOTTOM) HAVE BEEN CONSIDERED.

Purpose

to solve #3556
before: start a non async chunk Qwen3-omni server with FlashInfer CUTLASS Unquantized MoE backend

vllm bench serve --omni \
  --host 127.0.0.1 \
  --port 8666 \
  --backend openai-chat-omni \
  --endpoint /v1/chat/completions \
  --dataset-name random \
  --num-prompts 4 \
  --max-concurrency 1 \
  --request-rate inf \
  --random-input-len 2500 \
  --random-output-len 900 \
  --ignore-eos \
  --percentile-metrics ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration

now: start a non async chunk Qwen3-omni server with Triton Unquantized MoE backend

TTFP change from 46525.73 to 29900.31 ms on same bench test.

Test Plan

Test Result

---

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. Please provide the test scripts & test commands. Please state the reasons if your codes don't require additional test scripts. For test file guidelines, please check the test style doc
• The test results. Please paste the results comparison before and after, or the e2e results.
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model. Please run mkdocs serve to sync the documentation editions to ./docs.
• (Optional) Release notes update. If your change is user-facing, please update the release notes draft.

BEFORE SUBMITTING, PLEASE READ https://github.com/vllm-project/vllm-omni/blob/main/CONTRIBUTING.md (anything written below this line will be removed by GitHub Actions)
@ZeldaHuang @amy-why-3459 @tzhouam @Gaohan123 @hsliuustc0106

[gcanlin]

cc @Isotr0py @ZJY0516

[amy-why-3459]

VLLM_USE_FLASHINFER_MOE_FP16 seems to default to 0.

# Allow use of FlashInfer BF16 MoE kernels for fused moe ops.
    "VLLM_USE_FLASHINFER_MOE_FP16": lambda: bool(
        int(os.getenv("VLLM_USE_FLASHINFER_MOE_FP16", "0"))
    ),

However, it will check whether the VLLM_USE_FLASHINFER_MOE_FP16 environment variable is set.

# Handle explicit FlashInfer FP16 configuration.
    if envs.is_set("VLLM_USE_FLASHINFER_MOE_FP16"):
        if not envs.VLLM_USE_FLASHINFER_MOE_FP16:
            if UnquantizedMoeBackend.FLASHINFER_TRTLLM in AVAILABLE_BACKENDS:
                AVAILABLE_BACKENDS.remove(UnquantizedMoeBackend.FLASHINFER_TRTLLM)
            if UnquantizedMoeBackend.FLASHINFER_CUTLASS in AVAILABLE_BACKENDS:
                AVAILABLE_BACKENDS.remove(UnquantizedMoeBackend.FLASHINFER_CUTLASS)

[gcanlin]

Would it happen regression for thinker-only when VLLM_USE_FLASHINFER_MOE_FP16=0?

[R2-Y]

Would it happen regression for thinker-only when VLLM_USE_FLASHINFER_MOE_FP16=0?

regression for both thinker & talker when VLLM_USE_FLASHINFER_MOE_FP16=1,
VLLM_USE_FLASHINFER_MOE_FP16 = 0 -> triton
VLLM_USE_FLASHINFER_MOE_FP16 = 1 -> flashinfer

[amy-why-3459]

Would it happen regression for thinker-only when VLLM_USE_FLASHINFER_MOE_FP16=0?

regression for both thinker & talker when VLLM_USE_FLASHINFER_MOE_FP16=1, VLLM_USE_FLASHINFER_MOE_FP16 = 0 -> triton VLLM_USE_FLASHINFER_MOE_FP16 = 1 -> flashinfer

Even if the environment variable VLLM_USE_FLASHINFER_MOE_FP16 is not set, the flash infer backend will still be used.

[ZJY0516]

Please add a todo in comment to remove this later. It's a vllm moe performance regression

[R2-Y]

Please add a todo in comment to remove this later. It's a vllm moe performance regression

DONE

[R2-Y]

reopen because regression issues still exist in the main branch, the environment setting is only introduced in CI.