the naming is confusing... these models are aiming to equal or beat LLaMa by reproducing the trainign data and methodology that was used for LLaMa

But the actual model architecture is slightly different, based on Pythia

I guess what is needed is a pythia.cpp https://github.com/ggerganov/llama.cpp/issues/742

No, llama.cpp only works with llama-based models, like base llama, alpaca, vicuna, ...