Google open-sources TorchTPU, a native PyTorch stack for TPUs co-developed with Meta: Transformers, vLLM and SGLang run on TPUs, and Google reports up to 52% MFU across 8,192 Ironwood chips

Why it matters

【Key points】

On Oct 20 (11:30 UTC), at PyTorch Conference 2026, Google Cloud open-sourced TorchTPU, its native PyTorch stack for TPUs. Google says TPUs now join GPUs and CPUs as a native PyTorch option, so existing PyTorch code runs on TPUs with minimal changes.

1. Co-developed with Meta's PyTorch team: debug in eager mode, then use torch.compile for production; standard PyTorch APIs and native distributed primitives.

2. Ecosystem: Hugging Face Transformers v5 deploys to TPUs with device_map="tpu", and Diffusers runs natively; native TorchTPU backends for vLLM (with Inferact) and SGLang (with RadixArk); native Ray support; a TPU backend upstreamed to VeRL (preview).

3. Google's own training numbers: TorchTitan validated on up to ~8,000 Ironwood TPUs (v7x), up to 52% MFU on Qwen 3 8B, near-linear scaling to 8,192 Ironwood chips.

4. Also announced: LiteRT (formerly TensorFlow Lite) joins the PyTorch Ecosystem via litert-torch, and a TPU agent suite in early preview on GitHub.

5. Meta's PyTorch engineering director says TorchTPU lets Meta deploy PyTorch workloads, including LLM, ads and recommendation use cases, to Google TPUs.

【Verifiable】

Google Cloud blog: https://cloud.google.com/blog/topics/ai-infrastructure/torchtpu-is-now-open-source-run-native-pytorch-on-tpus/

GitHub: https://github.com/google-pytorch/torch_tpu

vLLM backend: https://github.com/vllm-project/vllm-torchtpu

SGLang backend: https://github.com/sgl-project/sglang-torchtpu

【Our read】

1. TPUs used to mean JAX or PyTorch/XLA. A native backend, with Transformers, vLLM and SGLang on board at launch, is the missing software piece for selling TPU capacity to PyTorch users.

2. Amazon open-sourced TorchNeuron for Trainium at the same PyTorch Conference; both clouds are making their own chips first-class PyTorch backends, a direct shot at CUDA's moat.

3. The MFU and scaling figures are Google's own; Meta gives no scale. Real migration cost will show with adoption.