At PyTorch Conference North America 2026 in San Jose on October 20, Google and Meta open-sourced TorchTPU, a native PyTorch backend for Google TPUs.

Bill Jia of Google Cloud and Meta software engineer Zongwei Zhou presented it on stage, and the SGLang project posted the same day that it was open-sourcing TorchTPU for native, hardware-accelerated PyTorch execution on Cloud TPUs.

Rather than one library, four repositories shipped: TorchTPU as the core backend, vLLM-TorchTPU and SGLang-TorchTPU for inference, and TorchTitan for large-scale training across dense, mixture-of-experts and diffusion workloads.

The torch-tpu package hit PyPI as version 0.2.0 on October 19, 2026, with Linux x86-64 wheels for Python 3.11–3.14, reported as Apache 2.0 licensed.

The surprise is how long this stayed quiet: Google started the effort in mid-2025 with Meta's PyTorch team, roughly a year before the public reveal, according to the PyTorch project's October 2026 account.

For years the TPU pitch for PyTorch teams sat behind a translation layer or a rewrite into JAX; a native backend that co-developed by the PyTorch team itself removes that tax.

Google says TorchTPU integrates with torch.compile, supports eager-execution optimizations and delivers native hardware acceleration — eager mode matters because TPU software historically leaned on ahead-of-time graph compilation that punishes debugging.

InfoQ reports the vLLM and SGLang inference engines run natively on the new backend, meaning teams can keep their serving setup and swap the hardware underneath.

The TorchTPU Team's launch post says that for bounded dynamism support on SymInts, the project evaluated correctness on 839 causal LLM models from the 5K corpus and confirmed correct numerics on 834 — 99%. Its CI/CD infrastructure is integrated into the official PyTorch CI HUD.

Google also reports Llama 3 405B reached 57% MFU on 1,024 TPU v7x chips, a vendor-reported figure rather than an independent benchmark.

The October 20 announcement describes TorchTPU as production-used for inference and training by some large Google Cloud customers while remaining in active development.

Caveats remain: no support matrix exists for TPU generations v5e, v5p, v6e and v7 (Ironwood), and a stated six-to-twelve-month roadmap targets production hardening, precompiled dynamic ATen kernels, XLA bounded dynamism and GKE packaging.

TorchTPU lowers the bar for open-weight models to reach a second kind of silicon: anyone pulling weights through Hugging Face Transformers and serving with vLLM or SGLang can now ask whether a model also runs well on TPUs, and Hugging Face's October 2026 post describes the project as adding 1 native TPU backend to PyTorch.

The open-source push is global momentum, not just a Google play — at the World Internet Conference in Wuzhen on November 3, Alibaba chairman Joe Tsai called open source the only workable route to a sustainable AI industry as demand outgrows computing power.

If accelerator supply stays tight, as Google's own leadership has signaled, a native PyTorch route to TPUs widens the pool of developers who can use the chips that exist without a JAX rewrite.