Google has launched TorchTPU, an open-source backend designed to enable native PyTorch execution on its Cloud TPU hardware. The project, announced on October 20, 2026, at the PyTorch Conference, was developed in collaboration with the Meta PyTorch team. Development of the backend reportedly began in mid-2025.
The system supports eager execution and approximately 450 ATen operations. It is designed to scale across thousands of chips, achieving 52% model FLOPs utilization during training of the Qwen 3 8B model on a cluster of 8,192 TPU v7x units. The software is compatible with Python 3.11 and later versions.
Distributed across four GitHub repositories under the Apache 2.0 license, TorchTPU allows for commercial use. It integrates with Hugging Face Transformers, enabling thousands of models to run with minimal code modifications. Developers can migrate existing workloads by changing the device call from .to(cuda) to .to(tpu).
Google positions TorchTPU as a production-ready replacement for PyTorch/XLA, which previously served as the primary method for running PyTorch on TPU hardware. The company reports that some Cloud customers have already adopted the new backend for their operations.
