It is fixed in this PR/branch: https://github.com/ggml-org/llama.cpp/pull/15181