TorchTPU

Description

This project provides a PyTorch backend for Google's Tensor Processing Units (TPUs), enabling users to run PyTorch models and operations on TPU devices. It includes custom ATen kernels for various PyTorch operations, a compilation cache for optimizing execution, and examples for distributed training and model inference.

Features

• PyTorch ATen kernel implementation for TPU.
• Compilation cache for optimized execution of XLA computations.
• Support for distributed training (Data Parallelism, Tensor Parallelism).
• Examples for various models (Llama, Qwen, ResNet, DLRM, minGPT, UNet).
• Integration with torch.compile() for graph mode execution.
• Utilities for debugging and benchmarking.

User Guide and Interactive Notebooks

A user guide is available at http://google-pytorch.github.io/torch_tpu/ . This is a static version of the interactive notebooks. They are available in the docs folder.

The notebooks are Marimo (similar to Jupyter). They can be run with marimo edit file.py and then you can access them in your browser. You will need to be running this from a TPU VM, a v6e GCE instance is easiest.

Installation

The same instructions are available in the user guide linked above.

First use python 3.12 virtual environment

python3.12 -m venv .venv
source .venv/bin/activate

or, with uv

uv venv .venv
source .venv/bin/activate

Install via pip

Authenticate first (for uv prepend uv to all pip commands):

pip install keyrings.google-artifactregistry-auth
gcloud auth login
gcloud auth application-default login

Install from latest nightly:

pip install --pre --index-url "https://oauth2accesstoken:$(gcloud auth print-access-token)@us-python.pkg.dev/ml-oss-artifacts-transient/torch-tpu-virtual-registry/simple/" torch_tpu

Build from source

Wheels can be built via:

bazel build --config=wheel_common //ci/wheel:torch_tpu_wheel

All OSS Bazel builds default to optimized mode (-c opt) automatically via .bazelrc. Unoptimized builds (-c fastbuild, -c dbg) can degrade performance and are not used for wheel builds. If you are developing or debugging C++ code and explicitly want an unoptimized wheel, pass --//:allow_unoptimized_wheel=True:

bazel build -c dbg --config=wheel_common //ci/wheel:torch_tpu_wheel \
    --//:allow_unoptimized_wheel=True

--config=wheel_common is required, and the build fails at build time without it. The config sets --//:wheel_build=True, which routes the shared XLA/MLIR backend through the torch_version-reset transition so that pywrap factors it into a single libxla_base.so; built without the flag, every per-version common ships its own copy of the backend and the resulting wheel aborts on import torch with duplicate static registrations.

wheel_common also turns on the Remote Build Execution (RBE) remote cache and remote executor. If you do not have RBE credentials, append --config=no_rbe to strip those flags back out:

bazel build --config=wheel_common //ci/wheel:torch_tpu_wheel --config=no_rbe

The order matters here: --config=no_rbe has to come after --config=wheel_common, because the later config is the one that wins for the remote-cache and remote-executor flags they both set.

Install wheel via:

cd <path_to_repo>
python3.12 -m venv .venv; source .venv/bin/activate
# The index-url includes the CPU version of torch at higher priority than the CUDA version.
pip install ../bazel-bin/ci/wheel/*.whl --index-url "https://oauth2accesstoken:$(gcloud auth print-access-token)@us-python.pkg.dev/ml-oss-artifacts-transient/torch-tpu-virtual-registry/simple/"

Note that this command will install the CUDA version of torch; to get the CPU-only build, use the --index-url parameter from the section on installing nightlies, or install the CPU build separately.

Getting Started

Some pointers to get you started:

• PyTorch: Official PyTorch documentation
• Tutorials: easy to understand PyTorch code using TorchTPU
• Examples: get you started with sample PyTorch models on TPU

Dependencies

The project uses Bazel for dependency management. Key Python dependencies are:

TBD

Bazel setup

This project uses Bazel for building and dependency management.

To install Bazel, we actually install Bazelisk, which is a transparent wrapper that handles downloading and installing Bazel itself.

There are several ways to install Bazelisk. Here are some common ways:

• If you have Go installed:

  go install github.com/bazelbuild/bazelisk@latest
  

• Download the binary directly:

  wget -O ~/bin/bazel https://github.com/bazelbuild/bazelisk/releases/download/v1.27.0/bazelisk-linux-amd64
  

See the Bazelisk README for more install instructions

However you install it, add it's location to PATH.

To verify the install, run bazel info in the repo's root directory:

bazel info

It should print information about the repo without errors.

VSCode clangd Setup

For C++ code navigation, it is recommended to use VSCode with clangd extension. Relying on a compilation database generated from Bazel action graph, clangd is able to enrich the editor with various smart features, including code completions, go-to definitions, etc.

1. Install VSCode extensions via UI or the following command:

   code --install-extension llvm-vs-code-extensions.vscode-clangd
   # Uninstall Microsoft C++ extensions to avoid interference.
   code --uninstall-extension ms-vscode.cpptools

2. Generate the initial compile_commands.json:

   ./setup_clangd.py

It may take a while (> 30 minutes) for the first run to make a full build, but reruns will be incremental and cheap. If you have run bazel build recently, generated files already exist so you can pass --no-build to skip waiting for heavy builds (note that on a fresh workspace, skipping the build may prevent clangd from resolving generated files). See more options in the script file-level comments.

Alternatively, you can invoke the script via nox (arguments after -- are forwarded):

nox -s refresh_compile_commands -- --no-build
nox -s refresh_compile_commands -- //src/torch_tpu/ops/...

Contributing

We welcome contributions to the PyTorch on TPU project! Please see our CONTRIBUTING.md for guidelines on how to contribute, including signing our Contributor License Agreement (CLA) and code review processes.

License

This project is licensed under the Apache 2.0 License. See the LICENSE file for details.