Amazon Opens TorchNeuron, a Native PyTorch Backend for Trainium Chips, as Apache 2.0 on Oct 20

Amazon and the PyTorch team at Meta announced TorchNeuron, a native PyTorch backend for AWS Trainium2 and Trainium3 that runs existing PyTorch code with a single `.to('neuron')` change — no rewrite needed. It ships Apache 2.0 at PyTorch Conference North America on Oct 20 and goes public beta the same day, with eager mode, torch.compile, FSDP2, tensor/pipeline parallelism, and FlashAttention supported natively. Deepgram reports Trainium processing ~60% more audio seconds per dollar than their GPU baseline; Cartesia, Hugging Face, Anthropic, Databricks, Rivian, Salesforce and poolside are listed as design partners.