AWS TorchNeuron Registers Trainium as Native PyTorch Backend

AWS announced TorchNeuron, an open-source native PyTorch backend for its Trainium chips, with a public beta set for Oct. 20 at PyTorch Conference North America¹. TorchNeuron registers Trainium through PyTorch's PrivateUse1 extension point, supporting standard APIs including torch.nn, torch.distributed, and torch.compile across Trainium2 and Trainium3. The backend supports distributed training and inference via FSDP2, tensor parallelism, pipeline parallelism, and DTensor, and ships Trainium-optimized operators including FlashAttention and MoE blockwise kernels through the open-source NKI-lib. The code will be released under Apache 2.0.