复现

[Gravity-vV]

请问如果只有40G的显存的显卡，有没有办法跑这个训练呢

[Yuancheng-Xu]

I think so, just use batch_size_per_device=1 and a higher gradient accumulation value.

[mattxu98]

On an RTX 4090 (24G), a line of bash poison_llava.sh --batch_size 512 --task_name Biden_base_Trump_target has been running for more than 24 hours and for at least the next 12 hours

[Yuancheng-Xu]

@mattxu98 I think the original post is asking about training the llava model. poison_llava.sh is for crafting the poison images.

I am surprised that 24G can handle a 512 batch size. Could you try to use a smaller batch_size (such as 10) and see whether it is still too slow?