Google released TorchTPU as open source, enabling PyTorch workloads to run on its TPUs with minimal code changes and familiar tools.
Google announced on October 20 that TorchTPU, its native PyTorch software stack for Tensor Processing Units, is available as an open-source project. The release was made at PyTorch Conference 2026 and co-developed with Meta’s PyTorch team. Google says the stack lets developers use standard PyTorch APIs to train, serve and deploy models on TPUs with minimal changes to existing workloads.
The launch connects TPU support with commonly used machine-learning tools and serving frameworks. Google lists vLLM, SGLang, llm-d, TorchTitan, Ray, TorchRec, Hugging Face and LiteRT among the supported ecosystem. Developers can use eager mode to debug models, then use torch.compile to optimize TPU utilization and reduce memory overhead, according to the company.
Google also described integrations with Hugging Face libraries. It says models in Transformers v5 can be deployed to TPUs by selecting the TPU device, while Diffusers supports execution modes and tensor parallelism across multiple accelerators. Google says it has set up nightly continuous testing of the Hugging Face Transformers suite on TPU infrastructure.
The release gives teams that already build with PyTorch a more direct route to testing Google’s accelerators, rather than requiring a complete rewrite of their applications. That may lower one software barrier to considering TPUs alongside GPUs and CPUs, though the announcement does not establish how easily every model or production workload will transfer or perform. Google also announced related serving integrations for vLLM and SGLang, developed with Inferact and RadixArk, respectively. The company’s post describes these tools as part of a broader effort to support open-source model serving on its hardware; developers can inspect the public TorchTPU repository and its implementation directly.