No need to convert models, 4bit LLaMA versions for GGML v2 available here:

https://huggingface.co/gotzmann/LLaMA-GGML-v2/tree/main