https://github.com/ggml-org/llama.cpp/discussions/15396