Google and Meta open-source TorchTPU, a native PyTorch backend for Google's TPUs
Models can now target TPUs with far fewer source changes than before, though getting good performance still takes tuning.
At PyTorch Conference North America 2026, Google and Meta open-sourced TorchTPU, a native PyTorch backend that lets models run on Google's TPUs without the intrusive code changes the older PyTorch/XLA bridge demanded. Meta, which created PyTorch, intends to move larger LLM workloads onto TPUs with it. Four repositories were released: TorchTPU, vLLM-TorchTPU, SGLang-TorchTPU and TorchTitan. Google's published benchmark reports a Qwen3 235B mixture-of-experts run across 8,192 TPU v7x chips reaching 28% model FLOPs utilization, and the sources caution that a one-line device change gets a model running but not tuned. Google also previewed a TPU agent suite on GitHub to help with onboarding, diagnostics and performance tuning. For developers, the practical takeaway is to budget for sharding and kernel work and to confirm TPU capacity before committing.
Why it matters
For people following AI infrastructure, the release means PyTorch models are no longer locked to GPUs by a fragile bridge layer, and Meta's stated plan to move larger LLM workloads onto TPUs suggests the hardware choice is becoming more open. The published 28% utilization figure on a very large mixture-of-experts run shows what is achievable at scale, while the caution that a one-line device change gets a model running but not tuned shows the gap that remains. Taken together, the news widens the set of accelerators available to PyTorch users without yet proving that TPUs match GPUs in everyday practice.
Matters or hype?
Does this story matter, or is it hype? Vote before you see what everyone else said.