Possible ways to reduce memory usage

[Niu0914]

Hello and thanks again for your open source work! Now I'm in 24 GiB server running on the 'multi - concept DreamBooth training', however my code in the 'self. The accelerator. The backward (loss)' is always error (cuda out of memory), I read the answer from issue3. I think I should be able to run, but I don't know what might go wrong. Can you give me some tips

[YuliangXiu]

Maybe you can refer to #10 (comment), and Diffuser GPU Memory, and DreamBooth Training to further reduce the GPU memory (even on 8GB).

[Niu0914]

Thank you, I can train dreambooth now！
However, when I was training main_mc, I found that the loss I printed was always 0. Why? (I did not change the source code given)