It's being worked on for upstream: https://github.com/ggml-org/llama.cpp/issues/29058