fix to adapt VLM multi img benchmarking with sglang.bench

[davilu-nvidia]

Overview:

benchmark CLI based on sglang.bench to allign withg sglang's PR
python -m sglang.bench_serving --model /raid/model_hub/Qwen/Qwen2.5-VL-7B-Instruct --num-prompts 1 --dataset-name image --random-input-len 128 --random-output-len 256 --image-count 8 --image-resolution 1080p --host localhost --port 8000 --backend vllm-chat --request-rate 1

Made some modifications to make this work, will init a PRI fixed qps =1 and only tune the total num_prompt to be 1 or 10

seems like --multimodal-encode-prefill-worker only works with llama4 but not qwen3 vl series due to vLLM limitations, so I put E and P processes on same GPU id 0 to demonstrate EP/D and limit P worker memory allocation to be 0.4 to avoid OOM.

Details:

sucess benchmark

Where should the reviewer start?

Related Issues: (use one of the action keywords Closes / Fixes / Resolves / Relates to)

• closes GitHub issue: #xxx

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.