Google has released TorchTPU, its native PyTorch stack for Google TPUs, as open source on GitHub. Announced at PyTorch Conference 2026, it lets developers train and serve models on TPUs through familiar PyTorch tools, according to the company.
Google co-developed TorchTPU with Meta’s PyTorch team. Google says existing workloads need minimal code changes. Developers can debug models in PyTorch’s eager mode, then use torch.compile to optimize them for TPU production workloads.
Google also worked with Hugging Face to add native TorchTPU support to its libraries. In Transformers v5, developers can select a TPU by setting device_map to tpu. Google says Diffusers supports eager mode, torch.compile and tensor parallelism on TPUs.
Read nextAmazon says TorchNeuron runs PyTorch code on Trainium without a rewriteServing tools add TPU support
TorchTPU works with the vLLM and SGLang serving engines. Google is also releasing llm-d 1.0 with Red Hat, IBM and a community of more than 500 contributors across 25 organizations. The software routes inference requests across model replicas and separates two stages of processing. On Google Kubernetes Engine, Google says teams can run it on either GPUs or TPUs through a common endpoint.
For training, Google has integrated TorchTPU with TorchTitan and added TPU support to Ray. It has also contributed a TPU backend and reference recipe to VeRL, whose TPU support is in preview. Google reports up to 52% model FLOPs utilization when training Qwen 3 8B through TorchTitan, with scaling demonstrated across up to 8,192 Ironwood TPUs.
LiteRT converts PyTorch models for devices
Google says LiteRT, formerly TensorFlow Lite, has joined the PyTorch ecosystem. LiteRT-Torch converts PyTorch models for use on Android, iOS, macOS, Windows, Linux, the web and IoT devices. Google also announced an early preview of a TPU agent suite on GitHub for onboarding, diagnostics, questions and inference tuning.
TorchTPU is available now on GitHub. Google says developers can use its Cloud documentation to start running PyTorch on TPUs today.
