why not directly use the original Qwen implementation on huggingface?

[FightingFighting]

Hi, can I ask why you dont directly use the original Qwen implementation on huggingface? could you give us a little bit hint about the difference between your implementation and original implementation ? so that we can implement Qwen3vl by ourself.

[LUOXIAO92]

Same question. I think we can just inherit the official implementation + insert the lora or maybe add a linear projector to downsample the embedding vector. Are there some special processes for fine tuning the models?