Google launches TorchTPU for native PyTorch on TPU

Google has launched TorchTPU, a backend for the PyTorch library, designed to run natively on Tensor Processing Units (TPUs) with full efficiency. This launch follows Amazon's release of TorchNeuron, which compiles PyTorch code to run on Trainium chips. TorchTPU is a collaborative development between Google and the PyTorch team at Meta, aimed at expanding PyTorch's usability to new chips beyond traditional Central Processing Units (CPUs) and Graphics Processing Units (GPUs).

Additionally, Google has partnered with Hugging Face to test the performance of various open models on TorchTPU, ensuring effective operation. The collaboration also extends to well-known libraries such as vLLM, SGLang, llm-d, TorchTitan, VeRL, Ray, and LiteRT to evaluate their usage on TPUs.