OptimalScale/LMFlow

is support llava model ?

Opened this issue · 3 comments

Thanks for your interest in LMFlow! LLaVA model has an architecture of ViT + Qformer + LLM, where LLM is normally freezed during training of LLaVA. If you would like a better LLM, you can use LMFlow with LISA to tune it in advance, upload the tuned model to huggingface, and then conduct the LLaVA training process.

Hope this information can be helpful 😄

@research4pan can support llava for yi-34B model?

Let us visit the engineering difficulty and we will get back to you once it is integrated 😄