Amazon introduces TorchNeuron backend for Trainium AI chips
The public beta starts on 20 October, but the initial version lacks support for some PyTorch features.
Amazon says its new TorchNeuron backend will let developers move existing PyTorch code onto Trainium AI chips by altering a single device setting rather than rewriting the code. Amazon engineers announced in a PyTorch blog post that public beta access begins on 20 October.
The beta launch is scheduled for the PyTorch Conference in San Jose. Trainium2 and Trainium3 are both supported, and the software will be open source under the Apache 2.0 license. TorchNeuron connects through PyTorch’s built-in hardware integration system, preserving the use of standard testing and scaling tools.
The engineers also outlined limitations: dynamic input shapes are not supported yet, and about 22,000 PyTorch tests are skipped. In a demonstration using identical training scripts on Trainium and a standard chip, Amazon said the model learned at an equal pace on the two.
Voice AI startups Deepgram and Cartesia are among the early users, along with the Lieber Institute for Brain Development, where TorchNeuron is used for drug research. Separately, Hugging Face has incorporated Trainium support into its AI tools, and Uber has agreed to use the chips through an expanded AWS deal.
