Amazon open-sources TorchNeuron so PyTorch code runs on Trainium without rewrites

Amazon announced TorchNeuron on October 19, 2026, a native PyTorch backend for its Trainium AI chips, and said it would release it as open source under the Apache 2.0 license at PyTorch Conference North America on October 20. The pitch is simple: developers keep their existing PyTorch code and change one setting to point it at Trainium instead of another chip.

TorchNeuron hooks into the standard way PyTorch lets new hardware register itself, so the usual testing, debugging and scaling tools are meant to work without changes. Amazon says that in a side-by-side demo the same training script learned at the same rate on a regular chip and on Trainium. It supports Trainium2 and Trainium3, works with TorchTitan for spreading training across many chips, and Hugging Face has added native Trainium support to its tools.

Amazon says it tests TorchNeuron against hundreds of thousands of models on every build and is sending changes upstream to make PyTorch less tied to particular chips. Early users named include voice AI company Cartesia and the Lieber Institute for Brain Development. Amazon also contributed to the vLLM inference project and open-sourced its performance kernels.

Why it matters

Porting code has been one of the biggest reasons teams stay on NVIDIA GPUs, so a drop-in PyTorch path makes Trainium a more realistic option for cutting AI compute costs.

Source: Amazon