Amazon and Google introduced new open-source tools at the PyTorch Conference to facilitate running PyTorch-based AI models on their proprietary hardware. These solutions aim to simplify the transition from Nvidia-optimized environments to Amazon Trainium and Google TPU accelerators.
Amazon released TorchNeuron, which enables running PyTorch projects on Trainium2 and Trainium3 chips without significant code modifications. The tool includes NKI language operators and a compiler for low-level model optimization, alongside support for the TorchTitan automation mechanism.
Google, in collaboration with Meta, launched the TorchTPU tool to allow existing PyTorch workloads to run on TPU accelerators with minimal code changes. It supports standard PyTorch APIs, distributed computing primitives, and advanced diagnostic tools.
Google also partnered with Hugging Face to ensure out-of-the-box support for open-weight models on TPU hardware. The ecosystem now integrates with industry-standard development tools including vLLM, SGLang, Ray, and TorchRec, reducing the need for extensive system re-engineering.
