DiffBIR icon indicating copy to clipboard operation
DiffBIR copied to clipboard

Vram needed when training

Open yimingz1218 opened this issue 2 years ago • 2 comments

Hello, how much is the minimum required vram for the stage2 training process? I train it on an A100 80G with batch size 16, image size 512*1024, and it nearly reaches the maximum. Are there some methods to make it more vram-saving?

yimingz1218 avatar Nov 08 '23 22:11 yimingz1218

你好,我设置batch size 1,image size 512*512,然后还安装了xformer,显存15g 还是显示不够,感觉按照你说的在a100上的内存占用情况,应该不至于跑不起来啊?

tqyunwuxin avatar Nov 24 '23 10:11 tqyunwuxin

Hello, I think the minimum vram required for a controlnet training (from scratch) with ideal batch size and image size is around 29 GB without xformer. 15 GB vram may not be enough for the model.

yimingz1218 avatar Nov 25 '23 01:11 yimingz1218