AWS has launched TorchNeuron in public beta, a native PyTorch engine that makes its Trainium2 and Trainium3 accelerators easier to use. Code can now be moved to a Trainium processor with a single line while retaining compatibility with standard PyTorch interfaces and executing operations immediately. The open-source solution, released on October 20, 2026, replaces a more complex PyTorch/XLA integration and accompanies version 2.33.0 of the Neuron SDK. AWS says it has more than 30 early users, including Cartesia, Deepgram and Hugging Face, while performance improvements reported by speech technology companies expand the scope for research applications, including drug discovery. TorchNeuron does not, however, support Trainium1 accelerators or the Inferentia chip family.
Source: Read the original article
