llama.cpp

52ff274b - ggml-cuda: avoid ROCm_Host compute on HIP integrated GPUs