The backend supports Trainium2 and Trainium3, eager execution, torch.compile, the Neuron Kernel Interface for custom kernels, and standard PyTorch distributed APIs. The announcement describes device selection through .to('neuron') and says code using torch.accelerator can run unmodified. The stated aim is to let developers use existing PyTorch models, libraries, and tools on Trainium without leaving the PyTorch ecosystem. The post does not establish parity in performance, model coverage, or migration effort across workloads. Source: https://pytorch.org/blog/introducing-torchneuron-native-pytorch-backend-for-aws-trainium
Introducing TorchNeuron: Native PyTorch backend for AWS Trainium (27 minute read)
In one minute
PyTorch announced TorchNeuron, an open-source native PyTorch backend for AWS Trainium, with public beta availability scheduled for October 20 under the Apache 2.0 license.
What changed: PyTorch announces TorchNeuron as its native, open-source Trainium backend, distinct from the prior Hugging Face backend release.