I really wish I could jump ship from Nvidia.

Having CUDA is just too important for work though, I really wish more ML libraries would add ROCm support. I understand it's now supported in Pytorch, but doing something like finetuning a Transformer model just isn't there yet (someone please, pleease tell me if I'm wrong)

The state of the closed source binaries required to get Nvidia GPU's running stable on Linux is a frustrating state of affairs.