Issue with Llama-3.2-11B-Vision-Instruct Supporting Only One Image Per Sample?

[iMaxb]

I am using Llama-3.2-11B-Vision-Instruct and the identity dataset. At the very beginning of training, it throws an error: ValueError: Mllama only supports one image per sample. When I use Llama-3.2-11B-Vision-Instruct directly for chat, it gives the same error. It only works properly when each message is accompanied by an image. Why is this happening?

[hiyouga]

Fixed in #6637