Google opens TorchTPU
On 20 October 2026, at PyTorch Conference 2026, Google announced that TorchTPU, its native PyTorch stack for Tensor Processing Units, is open source on GitHub. It lets existing PyTorch code run on TPUs with minimal changes; alongside it Google added TPU support to vLLM, SGLang, llm-d 1.0, TorchTitan, VeRL, Ray and Hugging Face. For training Qwen 3 8B it reports up to 52% model FLOPs utilisation (MFU) across up to 8,192 Ironwood chips.
Why it matters
One barrier to moving from GPUs to TPUs is the need to adapt code to another software stack; Google partly removes it by making TPUs a native option for PyTorch beside GPUs and CPUs. Whether developers take it up depends also on prices and the availability of chips, and the efficiency figure is Google's own measurement.
What the post says. The authors are Mark Lohmeyer and Bill Jia; the post announces TorchTPU as "our native PyTorch stack for Google TPUs", available as open source on GitHub and co-developed with the PyTorch team at Meta. A developer can debug a model in eager mode and then move to production performance through torch.compile. With Hugging Face, popular models from the Hub run on TorchTPU: in Transformers v5 deploying to TPUs is a matter of device_map="tpu", and the Diffusers library runs on TPUs natively. For inference: vLLM (with Inferact) and SGLang (with RadixArk) backends and llm-d 1.0 (with Red Hat and IBM, more than 500 contributors and 25 organisations, a CNCF Sandbox project). For training: TorchTitan, validated on up to 8000 Ironwood (TPU v7x) chips, where TorchTPU showed "up to 52% Model Flops Utilization (MFU) on Qwen 3 8B" with near-linear scaling across up to 8,192 Ironwood TPUs; a TPU backend for VeRL (preview) and Ray support. LiteRT (formerly TensorFlow Lite) joins the PyTorch ecosystem, and a TPU agent suite is in early preview on GitHub. In a quotation Meta calls TorchTPU a large step because it does not want to depend on a single hardware architecture. What the record does not claim. The TorchTPU repository itself was not opened; the 52% is Google's measurement on one model and one workload and the text does not compare it with GPUs. The post says nothing about prices or the availability of TPUs.