> I don't see the distinction Performant Pytorch compatibility can be achieved without CUDA compatibility. > In some cases one can replace rocm/openvino incom

> I don't see the distinction Performant Pytorch compatibility can be achieved without CUDA compatibility. > In some cases one can replace rocm/openvino incompatible libraries with pure PyTorch This is the crux of my point: AMD ought to focus on fixing these Pytorch incompatibilities instead of splitting focus with chasing CUDA translations.