[New Model]: Qwen2.5-VL

[yumlevi]

The model to consider.

new model family dropped

https://huggingface.co/collections/Qwen/qwen25-vl-6795ffac22b334a837c0f9a5

The closest model vllm already supports.

https://huggingface.co/collections/Qwen/qwen2-vl-66cee7455501d7126940800d

What's your difficulty of supporting the model you want?

No response

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[DarkLight1337]

@fyabc is your team planning to open a PR for this?

[zifeitong]

modular_transformers makes it a bit difficult to see the difference between Qwen2-VL and Qwen2.5-VL.

Here is a diff of transformers implementation with some reshuffling, modeling part only:
https://gist.github.com/zifeitong/1028fca68bbe4be45d2000c6dd16c157

The differences in processing, image_processing and config, is minimal.

[CuriousCat-7]

So, will it be supported in the future?

[DarkLight1337]

It is already supported in vLLM, please update your vLLM version.