Replaces --load-in-4bit with a more flexible --llama-bits arg to allow for 2 and 3 bit models as well. This commit also fixes a loading issue with .pt files which are not in the root of the models folder