Google open-sourced TorchTPU, enabling native PyTorch training and inference on its Tensor Processing Unit accelerators with minimal code changes.

TorchTPU was released as open source at PyTorch Conference 2026. Google says the stack was co-developed with Meta’s PyTorch team and is available on GitHub. It enables existing PyTorch workloads to run on Tensor Processing Unit (TPU) accelerators through familiar interfaces with minimal code changes.

The stack supports both interactive development and compiled execution. Developers can debug using eager mode and use `torch.compile` for production optimization. Google also lists standard PyTorch interfaces, advanced diagnostics, and native distributed primitives, alongside integrations with vLLM, SGLang, llm-d, TorchTitan, Ray, TorchRec, Hugging Face, and LiteRT.

Hugging Face support includes model deployment and multi-accelerator execution. In Transformers v5, developers can specify `device_map="tpu"`; Diffusers supports eager mode, `torch.compile`, and tensor parallelism across multiple accelerators. Google also describes dedicated continuous integration and continuous delivery pipelines that test the Hugging Face Transformers suite nightly on TPU infrastructure.

Meta described its involvement in terms of hardware diversity and production workloads. Its PyTorch engineering director cited large language models, advertising, and recommendation systems as relevant use cases. The supplied excerpt provides no numerical training or inference benchmark, so it does not establish a quantified advantage over CPUs or GPUs.