Google open-sources TorchTPU to run PyTorch on its AI chips — TechCrunch

Google announced the open-sourcing of the TorchTPU project, which enables developers to run PyTorch-based applications on its Tensor Processing Units (TPUs). To migrate to TPUs and the Google Cloud platform, it is enough to change the configuration without rewriting the application, TechCrunch reports.

An alternative to Nvidia graphics processors

PyTorch is a popular open-source toolkit for building, training and running artificial intelligence models. According to some estimates, it is used in 85% of AI applications. Meta initiated the development of PyTorch in 2016 and transferred the project to the Linux Foundation in 2022.

Many PyTorch features were initially designed for Nvidia graphics processors, which were already standard hardware for artificial intelligence research at that time. TorchTPU is intended to simplify the use of Google’s alternative chips for programs that were previously geared toward Nvidia GPUs.

More current news is available on the UA.News Telegram channel Telegram.

Amazon also released its own tool

A day earlier, Amazon introduced TorchNeuron, a tool for migrating PyTorch applications to its own Trainium AI chips and the AWS cloud without rewriting code. The company also said it would contribute code to the core PyTorch project that separates certain functions from Nvidia chips and makes them more accessible for different hardware.

Google reported cooperation with Meta, Hugging Face, Inferact, the developer of vLLM, RadixArk, the developer of SGLang, and other market participants to ensure TorchTPU’s compatibility with popular development tools. The company also launched the TPU agent suite for migrating applications and fixing errors, and adapted the LiteRT tool for better operation with PyTorch AI applications on phones, laptops and embedded devices.