3/ Extremely low memory footprint. LoRA trainable parameters are only 0.2% of total amount 
of model parameters.

→ View original post on X — @predibase,