Just stuff I do for the guys on /g/lmg. I will only post models here without any json files etc... You can just get those from my normal LLaMA-7-65B converts. Yes these work on cuda as well.

Currently for llama-30b-sft-oa-alpaca-epoch-2-4bit-128g The config.json's "vocab_size": 32000 needs to be "vocab_size": 32016 to be able to load it.