Google and Meta open-source TorchTPU, a native PyTorch backend for Cloud TPUs
Article timeline
The articles in this story. Times are ET.
- PyTorch BlogOfficialPickGoogle and Meta introduce TorchTPU, a native PyTorch backend for Google TPUs
AIGoogle and Meta announce TorchTPU, an open source project that runs PyTorch natively on Google TPUs. The backend supports eager execution and torch.compile, and the team reports validating out-of-box execution on about 5,000 Hugging Face models. Porting a model requires changing model.to("cuda") to model.to("tpu"), though peak performance needs further tuning.
- MarkTechPostNewsPickGoogle open-sources TorchTPU, a native PyTorch backend for Cloud TPUs
AIGoogle has open-sourced TorchTPU, a native PyTorch backend for Cloud TPUs, co-developed with Meta and announced at PyTorch Conference 2026. Google says it replaces PyTorch/XLA and recommends it for all new workloads. Self-reported results include 52% MFU training Qwen 3 8B with TorchTitan on Ironwood TPUs.
Swayam SinghXGoogle releases TorchTPU, open-source PyTorch support for TPUsAIGoogle has open-sourced TorchTPU, which lets native PyTorch code run on TPUs, according to a post linking to a Google Cloud blog announcement. The post's author says they are excited about the release, but it gives no further technical details.
- Google Developers BlogOfficialPickGoogle explains how to integrate TorchTPU into TorchTitan and optimize models for TPU
AIGoogle Developers describes running PyTorch training on Cloud TPUs through TorchTitan, using a native torch.device("tpu") backend and a fork of Meta's TorchTitan. The post lays out a model optimization playbook and reports results for Llama 3 405B (about 57% MFU on up to 1,024 chips), Qwen3 235B (32.42% MFU on a 1,024-chip slice), and FLUX.1, where a step time fell from 11.49 seconds to 2.78 seconds.