Amazon Web Services launches TorchNeuron for Trainium chips

Amazon Web Services (AWS) has released TorchNeuron, a native PyTorch backend designed for its Trainium2 and Trainium3 accelerators. The software, which entered public beta on October 19, 2026, and was subsequently open-sourced, aims to simplify the experimental research and training phases of AI development.

Unlike previous setups that relied on PyTorch/XLA integration—which required developers to manually insert specific calls into training loops—TorchNeuron allows for easier integration. Developers can move models to Trainium hardware using a single line of code, maintaining compatibility with standard PyTorch APIs and supporting eager execution.

The release also includes version 2.33.0 of the Neuron SDK, featuring the Neuron Observer for hardware health monitoring, improved NKI functionality for low-level customization, and updated vLLM plugins. While the new backend supports the latest Trainium generations, it does not support older Trainium1 or Inferentia chip families.