[Script] Update new VLM accuracy test for Qwen3-VL

[qichu-yun]

Update new VLM accuracy test for Qwen3-VL

Inspired by sgl-project#15205

Start evaluating:

    export OPENAI_API_KEY=EMPTY
    export OPENAI_API_BASE=http://localhost:9000/v1
    export PYTHONPATH=/the/path/to/your/sglang/python

    python3 -m lmms_eval \
        --model=openai_compatible \
        --model_args model_version=/mnt/raid0/models/Qwen3-VL-235B-A22B-Instruct-FP8-dynamic/ \
        --tasks mmmu_val   \
        --batch_size 16 \

The result of TP8 on MI308 should be:

    openai_compatible (model_version=/mnt/raid0/models/Qwen3-VL-235B-A22B-Instruct-FP8-dynamic/), gen_kwargs: (), limit: None, num_fewshot: None, batch_size: 16
    | Tasks  |Version|Filter|n-shot| Metric |   |Value |   |Stderr|
    |--------|------:|------|-----:|--------|---|-----:|---|------|
    |mmmu_val|      0|none  |     0|mmmu_acc|↑  |0.607 |±  |   N/A|

[sammysun0711]

May I know why mmmu accuracy measured with lmms_eval: 0.607 is different from accuracy 0.584 measure with benchmark/mmmu/README.md in CI: https://github.com/zejunchen-zejun/sglang/actions/runs/20593023968/job/59141786653

[qichu-yun]

May I know why mmmu accuracy measured with lmms_eval: 0.607 is different from accuracy 0.584 measure with benchmark/mmmu/README.md in CI: https://github.com/zejunchen-zejun/sglang/actions/runs/20593023968/job/59141786653

Because accuracy often fluctuates within a certain range, scores between 0.57 and 0.61 are all considered acceptable. When I use the method with benchmark/mmmu/README.md, I often get different results.