I don't get why they didn't use llama.cpp instead.