Google launches TorchTPU to run PyTorch apps on TPUs
The open-source project lets developers move existing PyTorch applications to TPUs and Google Cloud with configuration changes instead of code rewrites.
Google released TorchTPU, an open-source project that lets existing PyTorch applications run on TPUs and on Google Cloud without rewriting code, according to digitaltoday.co.kr. Developers move their apps to TPUs with simple configuration changes.
Alongside the release, Google published a "TPU Agent Suite" to support application migration and bug fixing, and updated its on-device AI runtime LiteRT for better integration with PyTorch apps.
The company said it worked with Meta, Hugging Face, Inferact (vLLM) and RadixArk (SGLang) on TorchTPU compatibility.
PyTorch was developed by Meta in 2016, and many of its features were designed for NVIDIA GPUs because those chips were the standard hardware for AI research at the time. In 2022, Meta handed the framework over to the non-profit Linux Foundation.
On October 19, Amazon released its own PyTorch version, TorchNeuron, which moves PyTorch app code to Amazon's Trainium AI chips and to AWS without rewriting. According to TechCrunch, Amazon is submitting code to the PyTorch project to decouple various features from NVIDIA chips.
Quick answers
What is TorchTPU?
It is an open-source Google project that lets existing PyTorch applications run on TPUs and Google Cloud without rewriting code.
Do I need to change my PyTorch code to use TPUs?
Google says applications can be moved to TPUs and Google Cloud with simple configuration changes.
Which organizations worked with Google on TorchTPU compatibility?
Google said it collaborated with Meta, Hugging Face, Inferact (vLLM) and RadixArk (SGLang).