TorchTPU
Open-source Google and Meta PyTorch backend for TPUs: eager and torch.compile modes, ~450 ATen ops, model.to('tpu').
Overview
TorchTPU, announced on the PyTorch blog on October 20, 2026 (PyTorch Conference North America), is an open-source collaboration between Google and Meta's PyTorch team that adds a native PyTorch backend for Google TPUs, replacing the lazy-tensor PyTorch/XLA approach. It supports eager mode (including a new DeferAndFuse mode) and torch.compile built on TorchDynamo, about 450 ATen ops, custom kernels via JAX/Pallas or Helion, and DTensor sharding. Google says it already runs training and inference in production for some of its largest Cloud customers and validated out-of-box execution of the 5,000 most downloaded Hugging Face models.
This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.
Key features
- Port with model.to('tpu') and dist backend 'tpu_dist'; no JAX dependency in the runtime
- Eager mode with DeferAndFuse plus torch.compile via TorchDynamo
- ~450 ATen ops with forward and backward for all Core ATen ops needed by major architectures
- Custom kernels via JAX/Pallas and the Helion DSL
- Validated with Hugging Face Transformers v5, Diffusers, vLLM, SGLang, TorchTitan, Ray, and TorchRec
- Bounded dynamism to avoid XLA recompilation on shape changes
Use cases
- Moving existing PyTorch training or serving code from GPUs to Cloud TPUs
- Running Hugging Face models on TPUs without JAX rewrites
Pricing
Open source (GitHub: google-pytorch/torch_tpu); license not stated in the announcement. TPU capacity is billed via Google Cloud.
Pricing reflects the source record and may change. Confirm current terms with the provider.
Dates and sources
- Released
- 20 Oct 2026
- Record updated
- 23 Oct 2026
- Record last checked
- 23 Oct 2026