Google open-sources TorchTPU for running PyTorch on TPUs
Original title: Google open-sourced TorchTPU, so PyTorch on TPUs just got a lot easier
TPU support also landed in vLLM, SGLang, TorchTitan, Ray, and Hugging Face libraries, with two pieces still in preview.
Google released TorchTPU as open source so PyTorch models can run on Tensor Processing Units using eager mode and torch.compile. Hugging Face Transformers v5 can send a model to a TPU, and Diffusers supports tensor parallelism. Meta says it already uses TorchTPU for LLM, ads, and recommendation workloads. Google reports up to 52% model FLOPs utilization on Qwen 3 8B with near-linear scaling to 8,192 Ironwood TPUs. VeRL’s TPU backend and a TPU agent suite remain in preview.
Unconfirmed: Utilization, scaling, and token-output figures are vendor or customer claims, not independently verified.
- Google says existing PyTorch GPU workloads need minimal code changes.
- Google runs the Transformers test suite on TPUs every night.
- Google released llm-d 1.0 with Red Hat and IBM.
- Zenteiq.ai reports output rising from 1.47 billion to 5.97 billion tokens per 1,000 TPU v6e chips.
- LiteRT, formerly TensorFlow Lite, was added to the PyTorch ecosystem.
- Vendors
- Organizations
- MetaHugging FaceRed HatIBMZenteiq.ai
- AI models
- Qwen 3 8B
From the original article
2 min read at helpnetsecurity.com
Google released TorchTPU as open source, giving PyTorch developers a native way to run models on the company’s Tensor Processing Units. Google also announced TPU support in the vLLM and SGLang serving engines, the TorchTitan training framework, Ray, and Hugging Face’s libraries. Teams that write their models in PyTorch and run them on GPUs by default have the most to gain.
By Anamarija Pogorelec · helpnetsecurity.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.