That also comes upstream from llama.cpp https://github.com/ggml-org/llama.cpp/discussions/4345