### What this PR does / why we need it? Add some of the pitfalls I ran into when using AISBench to test multi-modal models. - vLLM version: v0.11.0 - vLLM main: https://github.com/vllm-project/vllm/commit/2918c1b49c88c29783c86f78d2c4221cb9622379 --------- Signed-off-by: gcanlin <canlinguosdu@gmail.com>