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?

[memray]

@FightingFighting @LUOXIAO92
Thanks for the question. I recall that we had to customized some HF models to fix some bugs for both training and inference (such as obtaining embeddings in certain format, incompatibility with GradCache). Feel free to use the official backbone implementation. Also welcome to submit PR for new backbones. Thanks!

[Shengnan-Zhu]

@FightingFighting @LUOXIAO92 Thanks for the question. I recall that we had to customized some HF models to fix some bugs for both training and inference (such as obtaining embeddings in certain format, incompatibility with GradCache). Feel free to use the official backbone implementation. Also welcome to submit PR for new backbones. Thanks!

Hi, @memray , I noticed that you have modified the forward() function. If i want to use Qwen3-VL as backbone, Is it necessary to implement the same modification?