As always, code is the best documentation: https://github.com/ggerganov/llama.cpp/blob/8dd1ec8b3ffbfa2d...