Mental Note

[d8ahazard]

Putting this here because I'm working on something else and don't want to forget...

In the train_dreambooth.py file, somewhere around line 1068:

for epoch in range(args.num_train_epochs): try: unet.train() if args.train_text_encoder and text_encoder is not None: text_encoder.train()

What happens to VRAM if we do unet.to("cpu") except or when we call unet.train(), and the same with the text_encoder. Do we need to keep them in GPU if not running train() on them?