Google and Meta open-source TorchTPU, a native PyTorch backend that runs models on TPUs with one line changed
lmoroney · x · 2026-10-21
Google and Meta announced TorchTPU at the PyTorch Conference: an open-source, native PyTorch backend for TPUs, started at Google in mid-2025, aiming to make porting as simple as changing .to("cuda") to .to("tpu").
Key points:
- Unlike the 2020 PyTorch/XLA (Lazy Tensor-based, requiring intrusive code changes, slow compiles, no eager mode) or TorchAX (shoehorning PyTorch into JAX), TorchTPU is built as a principled backend with eager execution as a top priority.
- Supports torch.compile, covers 450 ATen ops, and existing PyTorch workloads run out of the box on TPUs.
- The team claims industry-competitive inference and training performance, still powered by the XLA compiler stack under the hood.
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