Google releases open-source TorchTPU to ease migration of PyTorch apps

Google announced that TorchTPU, its version of PyTorch, is now available as an open-source project. The tool lets developers run PyTorch-based AI applications on Google’s Tensor Processing Units (TPUs) and move them to Google Cloud with a simple configuration change, without rewriting the code. PyTorch is an open-source toolkit used to build, train and run AI models, while many of its features were developed to work with Nvidia GPUs.

Amazon announced TorchNeuron on Monday, offering a similar migration path to its Trainium chip and AWS, also without rewriting applications. Amazon also plans to submit code to the main PyTorch project to make its features less dependent on Nvidia hardware. Google says it worked with Meta, Hugging Face, Inferact and RadixArk to make TorchTPU compatible with popular tools, and launched an agent suite to help developers migrate applications and fix errors.