llama.cpp

28ff0958 - vulkan: use CPU writes in ggml_backend_vk_cpy_tensor_async if the context is idle (#28618)