Comment by wills_forward

So this could universally decrease the memory requirements by un-quantitized LLMs by 30%? Seems big if true.