Qwen2.5-VL Serverless Endpoint on RunPod

General
1

I’m trying to set up a Serverless Endpoint on RunPod with vLLM (with Qwen2.5-VL-3B-Instruct).
My goal is to get a lot of images descriptions (batch inference).

Here is how i set it up:

Docker Image:
runpod/worker-v1-vllm:v2.7.0stable-cuda12.1.0

GPU:
48GB pro
(L40, L40S, or RTX 6000 ADA)

ENV vars:

MODEL_NAME=Qwen/Qwen2.5-VL-3B-Instruct
DOWNLOAD_DIR=/runpod-volume
DTYPE=float16
GPU_MEMORY_UTILIZATION=0.90
ENABLE_PREFIX_CACHING=1
QUANTIZATION=bitsandbytes
LIMIT_MM_PER_PROMPT=image=10,video=0

This call with one image works :

curl https://api.runpod.ai/v2/xxxxxxxx/openai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer rpa_xxxxxxxxxxxxxxxx" \
  -d '{
    "model": "Qwen/Qwen2.5-VL-3B-Instruct",
    "messages": [
      {
        "role": "user",
        "content": [
          {
            "type": "text",
            "text": "Describe this image."
          },
          {
            "type": "image_url",
            "image_url": {
              "url": "https://www.site.com/images.jpeg"
            }
          }
        ]
      }
    ],
    "max_tokens": 1000,
    "temperature": 0.7
  }'

Now I have several questions.
Is it worth passing multiple images to the model in a single call? Will it be more efficient?
If so, how should I pass the parameters?
Did I miss anything in the ENV vars that would be important to go faster?
Thank you very much for any help or tips you can give me.