how to reduce video memory when running lora

[zhangnju]

HI

when I run qwen-image lora using this script, https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/lora/Qwen-Image-2512.sh, how do I reduce the video memory requirement?

I would like to have a try on the consumer GPUs, so the video memory is 32G or 48G. could you help share some solutions? I have tried to reduce the number of rank to be 16, but still OOM

Thanks

[Feng0w0]

Due to the current Qwen-Image not supporting zero3, GPUs with lower VRAM may cannot be trained even with Lora. You can try running the inference script to see if it can run properly.

[zhangnju]

hi @Feng0w0
Thanks for your reply. do other models of DiffSynth-Studio support zero3 or other CPU-offload? thus I can also try other models which have supported it.

Thanks

[Feng0w0]

To my knowledge, none of the models have been attempted on Zero3 so far. Models from the wan series or z-image series should be able to run.