Google and Meta launch TorchTPU, an open-source PyTorch backend that ports models to TPUs with one line
lmoroney · x · 2026-10-21
At the PyTorch Conference, Google and Meta announced TorchTPU, an open-source PyTorch backend for TPUs that started at Google in mid-2025. Porting is designed to be as simple as changing .to("cuda") to .to("tpu").
Key details:
- Supports eager mode and torch.compile
- Covers 450 ATen ops
- Team claims out-of-the-box validation on the 5,000 most-downloaded Hugging Face models
- Works with vLLM, SGLang, TorchTitan, and Ray
The blog is upfront that peak TPU performance may still require changes to shapes and sharding. Suggested first step: flip that one line on a model you already train on GPUs, run your usual eval, and compare before tuning anything.
More from Infra
- pi ships rec inference engine, runs locally on M5 Max in two lines — _rockt · 2026-10-22
- Researcher warns GitLab flaw is 'really bad,' urges users to patch now — basedjensen · 2026-10-22
- Meta Beats With $64.61B Revenue, Raises 2026 CapEx Guidance to $135-145B — RihardJarc · 2026-10-22
- PyTorch Con Demo Theater: 13 talks including Wan 2.2 video generation sped up 53.6x — PyTorch · 2026-10-22
- TorchTitan adds 5D parallelism for DeepSeek-V4, Qwen3.5, Kimi K3; Muon optimizer 79% faster — PyTorch · 2026-10-22
- JNTC cuts TGV glass substrate copper filling from a day to minutes, targets 2027 mass production — toptickcrypto · 2026-10-22