The project began at Google in mid-2025 and was announced on October 20, 2026. Its stated migration path is a minimal device change, such as replacing model.to("cuda") with model.to("tpu").
TorchTPU supports eager and compiled execution, including a DeferAndFuse eager mode, and integrates torch.compile through Torch Dynamo. The team says it supports about 450 Aten operations and TPU collectives for PyTorch distributed patterns including DDP, FSDP, and DTensor. It also reports validation on the 5,000 most downloaded Hugging Face models and integrations with vLLM, SGLang, TorchTitan, Ray, and TorchRec.
The source says TorchTPU is in active development and is used in production for inference and training by some large Google Cloud customers. It does not establish general availability terms, pricing, or that an unchanged GPU-oriented workload will achieve peak TPU performance; the authors explicitly say shapes, sharding, and operations may need tuning.