Can we use more than one image for training or inference of qwen-2vl?

[abhadu]

I am looking to train or inference qwen-2vl using llamafactory but its api only supports one image per prompt so is there any way to use more than one image for training or inference of qwen-2vl?

[hiyouga]

You can refer to this instruction for training with multiple images https://github.com/QwenLM/Qwen2-VL?tab=readme-ov-file#data-preparation
For inference, you can use the scripts in https://github.com/QwenLM/Qwen2-VL after merging the lora weights or using finetuned checkpoints

[abhadu]

but I'm having error "The number of images does not match the number of tokens" when I have only single prompt as I include multiple prompts equal to images the error gets resolved.

I'm training it using llamafactory.

[abhadu]

will the solution be using "image" token equal to images in single prompt?

[hiyouga]

sure

[abhadu]

Ok thanks, and is there any way to use the same in inference while using llamafactory chat interface as it only accepts one image?

[hiyouga]

nope for now, we recommend using the official one