That might be a philosophical problem that never prevented me from training models on Linux. The half-baked half-crashing AMD solutions just lead to wasting time I can spend on ML research instead.