Google Unveils TorchTPU for Native PyTorch Execution on TPU

Key summary

Google has launched TorchTPU, a backend library designed to run PyTorch code natively on Google's Tensor Processing Units (TPUs) with full efficiency. Developed in collaboration with the Meta PyTorch team, the project extends PyTorch support beyond traditional CPUs and GPUs to include TPU hardware. The initiative also involves partnerships with Hugging Face to test open-source models and collaborations with various libraries including vLLM, SGLang, llm-d, TorchTitan, VeRL, Ray, and LiteRT to validate performance on TPUs. Development of TorchTPU began in 2025, and the project was released as fully open source on GitHub during the PyTorch Conference 2026.