Amazon Opens TorchNeuron, a Native PyTorch Backend for Trainium, in Public Beta
AWS announced TorchNeuron, an Apache 2.0 open-source native PyTorch backend for Trainium2 and Trainium3, entering public beta on Oct. 20. Models move with .to('neuron'). Reuters via Business Standard says more than 30 early customers took part. Customer speed-ups are self-reported.
Nvidia's lock-in is software as much as silicon; Amazon is trying to shrink the rewrite cost of leaving it. AWS introduced TorchNeuron, an open-source (Apache 2.0) native PyTorch backend for Trainium2 and Trainium3, in a PyTorch Foundation blog ahead of public beta on Tuesday, October 20, with code at github.com/aws-neuron/torch-neuronx. Models move with .to('neuron') and can run in eager mode, an adaptive eager mode, or under torch.compile.
Reuters, via Business Standard, says the aim is to let PyTorch developers move workloads to Trainium with a configuration change rather than a rewrite, and that more than 30 early customers took part. The PyTorch post quotes Deepgram's CTO saying that in benchmarks it evaluated, Trainium processed about 60% more input audio seconds per dollar than its GPU baselines — a customer self-report.
Limits
- This is a public beta. Customer performance claims are company-reported. The Reuters copy names an Amazon spokesperson only by surname in the syndication we read, so that quote is omitted.