Google and Meta open-source TorchTPU, a native PyTorch backend for TPUs

Google announced TorchTPU as open source at PyTorch Conference 2026 on October 20, 2026. The native PyTorch backend for Cloud TPUs was co-developed with the PyTorch team at Meta and lets existing PyTorch code target a TPU through torch.device("tpu"), so TPUs sit beside GPUs and CPUs as a native PyTorch device. Google says TorchTPU replaces PyTorch/XLA and recommends it for all new workloads. The release spans four Apache 2.0 repositories covering the backend plus vLLM, SGLang and TorchTitan integrations, tested from TPU v5e up to 8,192 Ironwood chips.

Google reports 52 percent model FLOPs utilization training Qwen 3 8B with TorchTitan on Ironwood, a self-reported figure, and says it continuously tests about 5,000 Hugging Face models in eager and compiled modes. The release follows AWS TorchNeuron, the open source PyTorch backend for Trainium announced a day earlier and already listed on this page, so the two largest non-Nvidia silicon owners shipped the same portability move 24 hours apart.

Sources: MarkTechPost; Unite.AI