Editorial illustration for Google open-sources TorchTPU, a native PyTorch backend 50% faster than Strict
Google Open-Sources TorchTPU, 50% Faster PyTorch Backend
Google and Meta used PyTorch Conference 2026, held October 20 in the usual fall slot for the event, to announce TorchTPU, a new PyTorch backend that runs models directly on Cloud TPUs. The release comes as four separate open repositories: TorchTPU itself, plus vLLM-TorchTPU, SGLang-TorchTPU, and TorchTitan. TorchTPU is already on PyPI at version 0.2.0, requires Python 3.11 or newer, and ships under an Apache 2.0 license.
The pitch from Google's team is blunt: TorchTPU replaces PyTorch/XLA, the older bridge between PyTorch and TPU hardware, and should be the default choice for any new project. Where PyTorch/XLA relied on lazy-tensor graphs and often forced developers to mark out compilable regions by hand, TorchTPU plugs into PyTorch's PrivateUse1 interface and treats tensors as ordinary PyTorch tensors, no subclasses, no wrappers. It also drops the SPMD-only restriction that limited PyTorch/XLA, adding support for divergent MPMD execution.
Existing code can now reach TPUs through a plain torch.device("tpu") call, putting TPUs on equal footing with GPUs and CPUs as native devices. Google says the backend has been tested across TPU generations, from v5e up through 8,192-chip Ironwood clusters.
TPUs now sit beside GPUs and CPUs as a native PyTorch device. Google team states that TorchTPU replaces PyTorch/XLA and recommends it for all new workloads.
Why this matters
For years, TPUs meant either rewriting models in JAX or wrestling with PyTorch/XLA's quirks. TorchTPU closes that gap by making TPUs a first-class torch.device, the same way CUDA GPUs have worked all along. That's a real shift for teams who wanted TPU economics but didn't want to maintain a second codebase.
The Meta collaboration matters too: when the team that ships PyTorch itself co-builds the backend, compatibility promises carry more weight than a third-party wrapper ever could. Google's explicit recommendation to drop PyTorch/XLA for new work signals this isn't a side project, it's meant to be the path forward.
Still, "50% faster than Strict" is Google's own benchmark, not an independent one, and torch.compile-based stacks have a history of edge cases that only surface at scale. Anyone training on TPUs now has a decision to make: migrate early and get ahead of the curve, or wait for the broader community to find the rough edges first. Worth watching how fast vLLM-Torch and the other three repos mature, since that's where production reliability actually gets tested.
Related Reading
Browse all Google news →
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source Box CEO Aaron Levie Says AI Will Drive "Biggest Cybersecurity Upgrade
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
Related Topic Google to End Free Access to Gemini AI for Non-Subscribers
LLMs & Generative AI
Related Topic NVIDIA and Google Cloud let developers scale AI from prototype to production
Market Trends
From Archives Startup Builds AI Model to Decipher Plants' Electrical Language
From Archives Cloudflare Adds Clairic System One Model to Workers AI for Data Extraction
From Archives Anthropic, Google block OpenCode API keys after hackers ran up user charges
LLMs & Generative AI
From Archives Four Gemini 4 Argon agents beat Minecraft without human help
LLMs & Generative AI
Common Questions Answered
What is TorchTPU and how does it improve upon PyTorch/XLA?
TorchTPU is a native PyTorch backend developed by Google and Meta that enables models to run directly on Cloud TPUs with 50% better performance than PyTorch/XLA. It replaces PyTorch/XLA as the recommended approach for all new workloads by making TPUs a first-class torch.device, eliminating the need to rewrite models in JAX or work around PyTorch/XLA's limitations.
What repositories and packages are included in the TorchTPU open-source release?
The TorchTPU release consists of four separate open repositories: TorchTPU itself, vLLM-TorchTPU, SGLang-TorchTPU, and TorchTitan. TorchTPU is available on PyPI at version 0.2.0 under an Apache 2.0 license and requires Python 3.11 or newer for installation.
Why is the Meta collaboration on TorchTPU significant for PyTorch users?
Meta's involvement is significant because they are the team that ships PyTorch itself, which means compatibility promises and support for TorchTPU carry substantially more weight than a third-party wrapper would provide. This partnership ensures that the backend receives proper maintenance and integration with the PyTorch ecosystem.
How does TorchTPU change the workflow for teams wanting to use TPU economics?
TorchTPU eliminates the need for teams to maintain separate codebases or rewrite models for different hardware by making TPUs work seamlessly with PyTorch, just like CUDA GPUs do. This allows teams to leverage TPU's cost advantages without the complexity of managing multiple implementations or learning JAX.
When was TorchTPU announced and what event was used for the announcement?
TorchTPU was announced by Google and Meta on October 20, 2026 at the PyTorch Conference 2026, which was held in its usual fall slot. The announcement introduced TorchTPU as a major advancement in making TPUs more accessible to the PyTorch community.
Further Reading
- Your Guide to Hardware Acceleration & Compute Infrastructure at PyTorch Conference North America 2026 - PyTorch
- Core PyTorch Sessions at PyTorch Conference North America 2026 - PyTorch
- vLLM Sessions at PyTorch Conference North America 2026 - PyTorch
- PyTorch on TPUs: TorchTPU, XLA and Cloud TPU Explained - Sourcetrail
- Cloud TPU 上的PyTorch - Google Cloud
