IREE Releasing

This file documents the extant release process that IREE uses. This process and the automation (such as it is) has grown over many years and is due for a refresh. However, in the interests of documenting what exists, we attempt to do so here.

Nightly Core Releases

IREE development is primarily driven via automated nightly release snapshots. These are scheduled automatically each day by the schedule_candidate_release.yml workflow, which selects a green commit from main (for non optional CI tasks), created a tag of the format candidate-{YYYYMMDD}.{BUILDNUM} and schedules automation to populate the release.

The build_package.yml workflow then runs jobs to do builds for all platforms and packages, finally triggering the validate_and_publish_release.yml workflow.

Release artifacts are uploaded as a GitHub pre release and an index of files is updated by periodic automation at https://iree.dev/pip-release-links.html.

Some debugging notes for this process are available here: https://iree.dev/developers/debugging/releases/.

Nightly Release Packages

A number of packages are produced automatically:

  • iree-dist-*.tar.xz (manylinux x86_64 and aarch64): Install image of the binaries and development assets needed to use or depend on the C/C++ parts of the project.
  • iree-compiler: Binary Python wheels
  • iree-runtime: Binary Python wheels
  • iree-tools-tf and iree-tools-tflite: Pure Python wheels

Linux Builds

Binary Linux packages are built using a custom manylinux based Docker image hosted here: https://github.com/iree-org/base-docker-images/pkgs/container/manylinux_x86_64 using isolated self-hosted runners (only used for building checked in code) of sufficient size for building large components and GitHub managed runners for smaller components. The project aims to target all non-EOL Python versions with Linux builds on x86_64 and aarch64.

Windows Builds

Windows builds are built using GitHub-hosted runners. Due to the cost, the project aims to target the most recent version of Python only while building version N-1 for the first year of the lifecycle of the next version.

Only the Python iree-compiler and iree-runtime packages are built for Windows.

The release is published even if the Windows build fails. When this happens, it is fixed forward for the next snapshot.

MacOS Builds

MacOS builds are performed using GitHub-hosted runners. Due to the cost, the project aims to target the most recent version of Python only while building version N-1 for the first year of the lifecycle of the next version.

Only the Python iree-compiler and iree-runtime packages are built for MacOS.

The release is published even if the MacOS build fails. When this happens, it is fixed forward for the next snapshot.

Retention

The project will keep pre-release tagged releases on its releases page for a minimum of 6 months. Releases older than this can be purged.

Distribution to Package Registries

The following package registry projects are managed as part of the IREE release process:

PyPI

Deprecated projects no longer updated:

Build Promotion

There are presently two build promotion processes documented:

The versioning scheme for iree-turbine is rooted on the then-current PyTorch released version, with optional date-based dev/pre-release suffixes (i.e. rcYYYYMMDD or devYYYYMMDD) or intra PyTorch releases (i.e. postVVVV).

This process is being trialed to correspond with the 2.3.0 release of PyTorch. In this scenario, the pinned nightly build of IREE is considered current and promoted as part of the Turbine release to PyPI (and the release is marked as not pre-release on the GitHub releases page).

Promotions are done roughly monthly or at need. The schedule is shifted to account for extra factors as needed.

In the future, we would like to adopt a real versioning scheme (beyond the nightly calver+build number scheme) and manage promotion and pinning of the core IREE dep more explicitly and in alignment with how downstreams are using it.

Steps to Promote

There are multiple release artifacts that are deployed from this project:

  • shark-turbine wheel (transitional while switching to iree-turbine)
  • iree-turbine wheel
  • iree-compiler wheels
  • iree-runtime wheels

Typically we deploy IREE compiler and runtime wheels along with a turbine release, effectively promoting a nightly.

Building Artifacts

Start with a clean clone of iree-turbine:

cd scratch
git clone git@github.com:iree-org/iree-turbine.git
cd iree-turbine

Build a pre-release:

./build_tools/build_release.py --core-version 2.3.0 --core-pre-version=rcYYYYMMDD

Build an official release:

./build_tools/build_release.py --core-version 2.3.0

This will download all deps, including wheels for all supported platforms and Python versions for iree-compiler and iree-runtime. All wheels will be placed in the wheelhouse/ directory.

Testing

TODO: Write a script for this.

python -m venv wheelhouse/test.venv
source wheelhouse/test.venv/bin/activate
pip install -f wheelhouse iree-turbine[testing]
# Temp: tests require torchvision.
pip install -f wheelhouse torchvision
pytest core/tests

Push

From the testing venv, verify that everything is sane:

pip freeze

Push IREE deps (if needed/updated):

twine upload wheelhouse/iree_compiler-* wheelhouse/iree_runtime-*

Push built wheels:

twine upload wheelhouse/iree_turbine-* wheelhouse/shark_turbine-*

Install from PyPI and Sanity Check

TODO: Script this

From the testing venv:

pip uninstall -y shark-turbine iree-turbine iree-compiler iree-runtime
pip install iree-turbine
pytest core/tests