That would make the kernels the PyTorch Foundations's problem and they would have to set up CI infrastructure around AMD GPUs to maintain these kernels. For whatever reason, AMD really wants to keep everything in-house even though that has been a losing strategy so far.