InfoQ 中文T2·Infra
78 pts
At the PyTorch Conference North America 2026, AWS introduced TorchNeuron, a native PyTorch backend that lets developers run existing PyTorch code on AWS Trainium chips by simply swapping .to('cuda') for .to('neuron'), with no rewrite or revalidation required. The backend natively supports distributed training and parallel APIs including DDP, FSDP2 and DTensor, has been validated against 100,000 popular models and 800+ architectures, and lets enterprises share hardware capacity between training and inference. It follows Google's TorchTPU, signaling a race among cloud vendors to own the PyTorch accelerator backend layer.