separate the Image/LLM part

[okideal]

Excellent work!

I noticed in the paper that the LLM part is fine-tuned based on Qwen2.5-VL. Is it possible to separate the weights of the image part and the LLM part, so that other mature frameworks (such as TensorRT and Ollama) can be used for inference later?

[AFeng-x]

Hi, our model is built upon a fine-tuned Qwen2.5LM (to be clear, not the Qwen2.5-VL weights). Currently, when loading the model weights, we specify two different paths for the vision encoder and the LLM part, which allows the LLM to be separated. Therefore, the LLM part is compatible with other inference frameworks.