Google open-sources TorchTPU for native PyTorch on TPUs | Open Source
At PyTorch Conference 2026 on October 20, 2026, Google open sourced TorchTPU on GitHub and added TPUs as a native PyTorch option alongside GPUs and CPUs.
Published on MyPrivateClaw
Oct 21, 2026, 4:22 PM UTC
Coverage date
Oct 20, 2026
Last updated
Oct 21, 2026, 4:22 PM UTC
News summary
Google open sourced TorchTPU, its native PyTorch inference and training stack for Google TPUs, on GitHub at PyTorch Conference 2026 on October 20, 2026, adding TPUs as a native PyTorch option alongside GPUs and CPUs. In a production integration via TorchTitan, TorchTPU demonstrated up to 52% Model Flops Utilization (MFU) on Qwen 3 8B, validated across up to 8,192 Ironwood TPUs (TPU v7x); the figure is vendor reported and not independently benchmarked. TorchTPU is co developed with the PyTorch team at Meta and supports eager mode debugging plus torch.compile for production level performance. The release lets developers train, serve, and deploy models on TPUs with vLLM, SGLang, llm d, TorchTitan, Ray, TorchRec, Hugging Face, and LiteRT. The open source TorchTPU project on GitHub is licensed under the Apache 2.0 License. Hugging Face Transformers v5 enables TPU deployment by passing device…