If anyone is interested in a more 'pure' or 'scratch' implementation, check out https://github.com/michaelnny/QLoRA-LLM. (author here) It also supports 4-bit quantized LoRA, using only PyTorch and bitsandbytes, without any other tools.