Google has published TorchTPU on GitHub as an open-source project. Announced on Tuesday, October 20, 2026 at PyTorch Conference in a Google Cloud post by Mark Lohmeyer and Bill Jia, the software is a native PyTorch stack for Tensor Processing Units, the custom accelerators Google uses for its own Gemini models. The pitch is simple: TPUs should behave like any other PyTorch device, next to GPUs and CPUs.
That has not been true until now. As Hugging Face explains in its own launch post, PyTorch users on TPUs previously had to go through PyTorch/XLA, which relied on lazily executed tensors and could trigger surprise graph recompilations, or switch to JAX. TorchTPU instead runs in eager mode for debugging and uses torch.compile when developers want production speed. Google says it was co-developed with the PyTorch engineering group at Meta, and Meta's Jana van Greunen said it lets the company run large language model, ads and recommendation workloads on TPUs.
The release arrives with a long list of integrations. In Hugging Face Transformers v5, a model can be placed on a TPU by setting device_map to "tpu", and Diffusers supports eager mode, compilation and tensor parallelism. Hugging Face says its separate Optimum TPU library is no longer needed for these libraries. For serving, Google built a vLLM backend with Inferact and an SGLang backend with RadixArk, and it ties both into llm-d, the distributed inference project that reached version 1.0 the same day.
On the training side, TorchTitan has TPU recipes for dense models such as Llama 3 and Qwen 3, mixture-of-experts models and diffusion transformers. Google reports that TorchTitan training on TorchTPU reached up to 52% model FLOPs utilization on Qwen 3 8B and scaled almost linearly on as many as 8,192 Ironwood TPUs. A TPU backend for the VeRL reinforcement learning framework is in preview, and Ray can now schedule pipeline stages across CPUs, GPUs and TPUs. Google also said LiteRT, its on-device runtime formerly called TensorFlow Lite, is joining the PyTorch ecosystem through a litert-torch converter, and it put a TPU agent suite for onboarding and tuning into preview.
The business context is the battle over Nvidia's software lock-in. TechCrunch notes that most AI applications are written for Nvidia GPUs by default, and that Amazon released a comparable PyTorch backend for its Trainium chips, TorchNeuron, one day earlier. By moving existing PyTorch code to TPUs with little rewriting, Google makes it easier for customers to try its cloud hardware without abandoning familiar tools.
Several questions remain open. The utilization and scaling numbers come from Google and have not been reproduced independently, and partners quoted in the announcement are commercial collaborators. Google did not publish a full compatibility matrix for PyTorch operators, and it is unclear how quickly custom CUDA kernels used by many teams can be replaced. Open-sourcing the stack lowers the barrier to switching, but cost per token on real workloads will decide whether developers actually move.