why not vulkan
that way you dont waste time implementing cpu,cuda,amd,metal,intel
Vulkan is welcome, though I'm not sure the extent of ggml's vulkan support. This started as a cuda programming class project, and I have an nvdia card, so that's why my work is on a cuda version. I will support and review vulcan related contributions to this project if people want to contribute vulcan support, but a working cuda version is where my own dev efforts are going for now.
isn't on the repo https://github.com/ggerganov/ggml/blob/master/src/ggml-vulkan.cpp
Note to those interested in contributing: This is a good issue to work on! I will help you get started if you want to work on a vulkan or AMD version.