Latest / Chips, compute and energy

LaunchComputeFrontierUSConfirmed

Google open-sources TorchTPU, a native PyTorch backend for its TPUs, with Meta

See how this story connects

Google released TorchTPU as open source on GitHub on 20 October 2026, at PyTorch Conference 2026 in San Jose. Co-developed with the PyTorch team at Meta and started in mid-2025, it runs PyTorch on Google’s Tensor Processing Units as a standard device: existing code moves by changing .to(“cuda”) to .to(“tpu”), with eager mode for debugging and torch.compile for production. Google also announced TorchTPU backends for the vLLM and SGLang serving engines, TorchTitan, VeRL and Ray, TPU as a device option in Hugging Face Transformers v5, llm-d 1.0 with Red Hat and IBM, and LiteRT joining the PyTorch ecosystem. Google says TorchTitan training on TorchTPU reached 52% model flops utilisation on Qwen 3 8B across up to 8,192 Ironwood (TPU v7x) chips. Amazon announced TorchNeuron, an Apache 2.0 PyTorch backend for Trainium2 and Trainium3, on 19 October ahead of the same conference. The TorchTPU repository is under the Apache 2.0 licence.

Why it matters

PyTorch code written for Nvidia GPUs can now run on Google and Amazon chips with small changes, which weakens the hold of Nvidia’s CUDA software on where models are trained and served.

When
20 Oct 2026
ID
WTA-N-2026-0350
Who
Google, Meta, Amazon Web Services, Hugging Face, Nvidia, PyTorch Foundation

Summary by wheresthe.ai, from Google Cloud and the reporting credited above. Checked against the source on 21 Oct 2026.

Source copy: download the page as saved on 21 Oct 2026 (PDF). Also held by the Internet Archive.

Earlier in this Storyline: What led hereGoogle announces TPU v5p with 8,960-chip pods and says Gemini was trained on TPUs

Cite this story

WTA-N-2026-0350: “Google open-sources TorchTPU, a native PyTorch backend for its TPUs, with Meta.” wheresthe.ai, 20 Oct 2026. https://wheresthe.ai/d/intl-google-open-sources-torchtpu-software-for-pytorch-on-it-2026-10/ (source: Google Cloud).

Spotted an error? Email [email protected].