Thank you for your amazing work! I more or less use it as a llama cpp replacement cause I honestly can’t figure out a good prompt structure so I get bad results

Thank you for your amazing work! I more or less use it as a llama cpp replacement cause I honestly can’t figure out a good prompt structure so I get bad results. On my M1 MBP, I use it for stuff that can’t go to GPT but can only use upto 13B. I also run a 34B model with CUBLAS enabled on my Linux server by modifying the go files mentioned in one of the issue tickets.