Google has open-sourced TorchTPU, a native PyTorch backend for its TPU AI accelerators, and the launch comes with day-one support from the Hugging Face ecosystem — a major step toward making TPUs a first-class option for PyTorch developers.
Announced at PyTorch Conference 2026, TorchTPU was co-developed with the PyTorch team at Meta. It lets teams run existing PyTorch workloads on TPUs with minimal code changes: debug in eager mode, then switch to torch.compile for production performance. Meta says it is already using TorchTPU on key production workloads, including LLMs, ads, and recommendation systems.
Hugging Face announced support the same day. In Transformers v5, deploying a model to a TPU is as simple as passing device_map="tpu"; the Diffusers library runs natively on TPUs with support for eager mode, torch.compile, and tensor parallelism. Hugging Face says its older Optimum TPU package is no longer needed for Transformers and Diffusers models, and reports that 96.5% of the Transformers tests that run on TPU now pass.
For serving, a native PyTorch vLLM backend developed with Inferact adds continuous batching, quantization, and multi-chip parallelism on TPU hardware, while an SGLang backend targets structured and agentic workloads. Google also shipped version 1.0 of llm-d, an open-source distributed inference stack for Kubernetes.
Builders exploring open models can compare assistants in our AI chatbots comparison, including HuggingChat and open image tools like Stable Diffusion.
Source: Hugging Face
