[coredog64]

AWS uses tricks to accelerate PyTorch with Inferentia/Trainium. Haven’t used it, but I have tried the equivalent for Apple silicon and rage quit after wasting half a day.