[Misc] Adding MMMU-Pro vision dataset to serving benchmark

[ywang96]

This PR adds the support for MMMU-Pro vision dataset to the serving benchmark. This dataset is image-token-heavy (single image with average resolution of 1700x1600) and comes with a generic short text prompt, as describe in the dataset model card:

Vision: In this subset, questions are embedded within screenshots or photos, and models must integrate visual and textual information to answer correctly. No separate text is fed into the model.

Example command to run the benchmark

vllm serve llava-hf/llava-v1.6-mistral-7b-hf

python3 benchmarks/benchmark_serving.py \
--dataset-path MMMU/MMMU_Pro \
--dataset-name hf \
--hf-subset "vision" \
--hf-split test \
--model llava-hf/llava-v1.6-mistral-7b-hf  \
--backend openai-chat \
--endpoint /v1/chat/completions \
--num-prompts 100

Co-authored-by: @heheda12345

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
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 do one of these:

• Add ready label to the PR
• Enable auto-merge.

🚀

[ywang96]

Note: Currently the serving benchmark does not correctly count the input image tokens, therefore the reported total token generation rate is not accurate.

[Isotr0py]

Overall LGTM! I left a comment about dataset sampling, PTAL!

[Isotr0py]

LGTM!