Google's TorchTPU Lands on GitHub at Version 0.2.0 With Its Dependencies Section Marked "TBD"

Google's TorchTPU, a PyTorch backend for its Tensor Processing Units, is now on GitHub as google-pytorch/torch_tpu with an Apache 2.0 licence, version 0.2.0 on PyPI and a README whose dependency section reads "TBD".

Google's Developers Blog had said on April 7 that a public repository, with documentation and tutorials, was on the 2026 roadmap. The repo's API record shows it was created on March 9, the month before that post, and pushed to again on October 20. It had 139 stars and 6 forks that day, against 256 open issues and pull requests, according to GitHub's API.

What the README says it ships

The README lists custom ATen kernels, a compilation cache, torch.compile() integration and examples for distributed training and inference. The examples cover Llama, Qwen, ResNet, DLRM, minGPT and UNet. It requires Python 3.11 or later and Bazel to build from source. It names one hardware target: a v6e TPU VM, which it calls "easiest" for the notebooks.

Item	What the README or repo says
Version	0.2.0 on PyPI
Licence	Apache 2.0, plus MIT for the minGPT example
Dependencies section	"TBD"
Operator coverage list	None given
Performance claims	None given

Install traps the README spells out

Installing without the CPU index URL pulls the CUDA build of torch, the README warns. The command it gives is pip install torch_tpu==0.2.0 --extra-index-url https://download.pytorch.org/whl/cpu. Wheel builds fail without --config=wheel_common, and a build without remote-execution credentials needs --config=no_rbe. It also says unoptimised Bazel builds "can degrade performance".

The licence line is more complicated than the badge

GitHub's API reports the licence as "Other" (NOASSERTION), not Apache-2.0. The LICENSE file explains it: Apache 2.0 covers the repo, but a separate MIT licence, copyright 2020 Andrej Karpathy, covers examples/mingpt/impl/. A licence scanner sees two licences and declines to pick one. The project's code is Apache-licensed; the one directory is not.

What the April post promised

The Google Developers Blog post describes three execution modes: Debug Eager, Strict Eager and Fused Eager. Its eight credited authors, including Claudio Basile and Lee Howes, wrote that Fused Eager "consistently delivers a 50% to 100+% performance increase over Strict Eager". That is a vendor figure and compares Google's own modes, not TPUs against Nvidia GPUs.

The post said TorchTPU compiles through XLA rather than Torch Inductor, supports DDP, FSDPv2 and DTensor, and may recompile when sequence lengths or batch sizes change. Lee Howes, whom The Stack called a Google engineering lead, said: "The TPU should be an obvious choice for any PyTorch user to target."

The 2026 roadmap also lists Helion kernel support, first-class dynamic shapes, deeper vLLM and TorchTitan integration, and "validated linear scaling up to full Pod-size infrastructure". The README as fetched mentions none of those items.

For the equivalent effort on Amazon's chips, see the site's report on TorchNeuron for Trainium. For a different open release, see Meta's Muse Spark Open.

The README gives no date for a release beyond 0.2.0.