Model trained in 8-bit with LoRA

Usage:

with torch.autocast(): # needed for best performance