In the deep learning world every major framework works on top of Cudnn which works on top of CUDA Pytorch, Tensorflow you name it.
https://github.com/pytorch/pytorch/issues/10657
That is the state of Pytorch support for AMD GPUs.
> Who’s writing bare CUDA though?
I do. Not everything you can do with a CUDA card is deep learning. In fact that's just one of many applications.
If AMD could provide a backend for the most popular frameworks then they could skip over the CUDA patent issue completely.
The real problem is that it seems like AMD’s not investing substantially in software teams to make it happen.