I'm still blown away that AMD hasn't made it their top priority. I've said this for years. If I was AMD I would spend billions upon billions if necessary to make a CUDA compatibility layer for AMD. It would certainly still pay off, and it almost certainly wouldn't cost that much.

They've been doing it all the time and it's called HIP. Nowadays it works pretty well on a few supported GPUs (CDNA 3 and RDNA 4).

AMD did hire someone to do this and IIRC he did, but they were afraid of Nvidia lawyers and he released it outside of the company?

Moving target, honestly just get PyTorch working fully (loads of stuff just doesn’t work on AMD hardware) and also make it work on all graphics cards from a certain generation. The matrix of support needed GFX cards, architectures and software together is quite astounding but still yes that should have at least that working and equivalent custom kernels.