Hacker News new | ask | show | jobs
by kashifr 1129 days ago
An efficient finetuning approach that reduces memory usage enough to finetune a 65B parameter model on a single 48GB GPU while preserving full 16-bit finetuning task performance!