New technique can reduce memory needs of LLMs by 50x

Technology