Does tgi support image resize for qwen2-vl pipeline?

[AHEADer]

trafficstars

System Info

I try to deploy a qwen2-vl fine-tuned model with tgi and vllm, and I've found some results between these two frameworks are different. Seems that tgi consume more tokens compared to vLLM. I checked TGI's code and seems there miss the image resize logic? For Qwen2-VL pipeline, we will resize the image based on two args max_pixels and min_pixels.

Information

• [x] Docker
• [ ] The CLI directly

Tasks

• [ ] An officially supported command
• [ ] My own modifications

Reproduction

Deploy a Qwen2-VL-7B model on the inference endpoint, and upload a large image will trigger an error that the input tokens are larger than 32768

### Expected behavior

The server will resize the image based on preprocessor_config.json(max_pixels and min_pixels) and make sure the image tokens will not be too many for a request.

[ashwani-bhat]

@AHEADer can you provide the docker command you are using with Qwen2-VL-7B?