Google launched TorchTPU to help developers run AI apps on its own chips instead of NVIDIA hardware. The tool simplifies moving existing PyTorch apps to Google Cloud with a simple configuration change. Google also started a TPU agent suite to help developers fix errors while switching to their cloud systems.
PyTorch, an open-source toolkit for building, training, and running AI models, is used by 85% of such apps.
Developed by Meta (formerly Facebook) in 2016, it was designed to work specifically with NVIDIA's GPUs.
In 2022, Meta handed over PyTorch to the Linux Foundation, a nonprofit organization that oversees several important open-source projects.
Google's TorchTPU allows AI apps to run on its Tensor Processing Units (TPUs). It also simplifies the process of moving existing PyTorch apps to TPUs and Google Cloud, requiring just a simple configuration change.
This is similar to Amazon's TorchNeuron, which lets AI apps be ported without rewriting them.
Google has collaborated with Meta, Hugging Face, Inferact (creator of vLLM), and RadixArk (creator of SGLang) to ensure TorchTPU's features and code libraries work seamlessly with these popular AI development tools.
The tech giant also launched a dedicated tool called the TPU agent suite, which helps developers transition their apps to Google's chips and cloud while fixing errors along the way.