Google has open-sourced TorchTPU, a native PyTorch backend that makes its TPU AI chips "just another device" for the world's most popular AI framework. It was announced at PyTorch Conference 2026 on Tuesday, Oct. 20, and co-developed with the PyTorch team at Meta, Google Cloud says. Developers can now train, serve and deploy models on TPUs with standard PyTorch code instead of porting to JAX or the older PyTorch/XLA bridge.

What's in the release

  • Eager and compiled modes: debug in eager mode, then speed up with torch.compile. The PyTorch blog says TorchTPU supports about 450 ATen ops, including all core ops needed for major deep-learning architectures.
  • Broad model coverage: out-of-the-box execution validated on the 5,000 most-downloaded Hugging Face models, including Qwen, DeepSeek, Gemma, Llama, Mistral and Whisper, per the PyTorch blog.
  • Serving and training stacks: integrations with vLLM, SGLang, llm-d, TorchTitan, Ray and TorchRec, Google says. Four repositories were open-sourced: TorchTPU, vLLM-TorchTPU, SGLang-TorchTPU and TorchTitan.
  • Custom kernels can be written in JAX/Pallas or the Helion DSL.

Hugging Face says loading a Transformers model on a TPU now takes device_map="tpu", and that its older Optimum TPU library is no longer needed. Of the Transformers tests that run on TPU, 96.5% pass so far; the team has filed 34 issues, 19 of them already fixed.

TorchTPU "is used in production for both inference and training workloads for some of the largest Google Cloud customers," the PyTorch blog says, with a 6- to 12-month roadmap focused on production hardening.

Why it matters

Nvidia's grip on AI training rests partly on software: most models are written in PyTorch and tuned for its GPUs. Making TPUs a first-class PyTorch target lowers the cost of switching chips, which matters as AI labs shop for compute. Our earlier coverage of Google Cloud's Vera Rubin plans shows Google is still buying Nvidia at scale too. Tesla, which trains FSD on Nvidia hardware and is designing its own AI5 chips, isn't mentioned in the release.