Google and Amazon Release Native PyTorch Backends for Non-Nvidia Hardware

Google and Amazon have both released native PyTorch backends that allow ML workloads to run on non-Nvidia hardware. Google open-sourced TorchTPU, a native PyTorch backend for TPUs that replaces the older PyTorch/XLA layer. PyTorch/XLA queued operations for a compiler, making TPU debugging behave differently from GPU workflows; TorchTPU runs operations eagerly for debugging, then switches to torch.compile for production. TorchTPU reached up to 52% of peak TPU compute when training Qwen 3 8B. Separately, Hugging Face announced TorchNeuron, a tool built in collaboration with AWS to reduce the complexity of deploying Hugging Face models on AWS Trainium accelerators.