Add --quick_release Option for Improved VRAM Management in ZipLoRA Saving

mkshing / ziplora-pytorch

Implementation of "ZipLoRA: Any Subject in Any Style by Effectively Merging LoRAs"

MIT License

481 stars 33 forks source link

Add --quick_release Option for Improved VRAM Management in ZipLoRA Saving #8

Closed jinnsp closed 7 months ago

jinnsp commented 7 months ago

There's unnecessary VRAM usage in the saving phase of ZipLoRA training. This commit makes training fit within 24GB VRAM (like on a 4090), allowing one to train without needing the xformers option.