Skip to content
This repository has been archived by the owner on Nov 14, 2020. It is now read-only.


Repository files navigation

DEPRECATED: PLEASE USE cibuildwheel instead!

cibuildwheel covers everything this could do, and more, for all CI platforms. We help maintain cibuildwheel now, and development effort is directed there instead of maintaining this repo. See our guide for details on using cibuildwheel!

Azure Wheel Helpers

This repository holds a collection of wheel helpers designed by the Scikit-HEP project to build Python Wheels on Azure DevOps. This is designed for packages that require building; if you have a pure-Python project, producing a universal wheel is trivial without this helper collection. This collection assumes some standard paths and procedures, though some of them can be customized.

Azure provides manual pipeline triggering and release pipelines, making it slighly better suited for this than GitHub Actions, though otherwise they are very similar.

Supported platforms and caveats

TLDR: Python 2.7, 3.6, 3.7, and 3.8 on all platforms, along with 3.5 on Linux.

System Arch Python versions
SDist (all) all any (non-binary distribution)
ManyLinux1 64 & 32-bit 2.7, 3.5, 3.6, 3.7, 3.8
ManyLinux2010 64-bit 2.7, 3.5, 3.6, 3.7, 3.8
macOS 10.9+ 64-bit 2.7, 3.6, 3.7, 3.8
Windows 64 & 32-bit 2.7, 3.6, 3.7, 3.8
  • Linux: Python 3.4 is not supported because Numpy does not support it either.
  • manylinux1: Optional support for GCC 9.1 using docker image; should work but can't be called directly other compiled extensions unless they do the same thing (think that's the main caveat). Supporting 32 bits because it's there for Numpy and PPA for now.
  • manylinux2010: Requires pip 10+ and a version of Linux newer than 2010. This is very new technology. 64-bit only. Eventually this will become the preferred (and then only) way to produce Linux wheels. Optional modern GCC image available.
  • MacOS: Uses the dedicated 64 bit 10.9+ builds. We are not supporting 3.5 because those no longer provide binaries (could use 32+64 fat 10.6+ but really force to 10.9+, but will not be added unless there is a need for it).
  • Windows: PyBind11 requires compilation with a newer copy of Visual Studio than Python 2.7's Visual Studio 2008; you need to have the [Visual Studio 2015 distributable][msvc2015] installed (the dll is included in 2017 and 2019, as well).


Azure does not recognize git submodules during the configure phase. Therefore, we are using git subtree instead.

This repository should reside in /.ci in your project. To add it:

git subtree add --prefix .ci/azure-wheel-helpers master --squash

You should make a copy of the template pipeline and make local edits:

cp .ci/azure-wheel-helpers/azure-pipeline-build.yml .ci/azure-pipeline-build.yml

Make sure you enable this path in Azure as the pipeline. See the post here for more details.

You must set the variables at the top of this file, and remove any configurations (like Windows) that you do not support:

  package_name: my_package    # This is the output name, - is replaced by _
  many_linux_base: "" # Could also be "skhep/manylinuxgcc-"
  dev_requirements_file: .ci/azure-wheel-helpers/empty-requirements.txt
  test_requirements_file: .ci/azure-wheel-helpers/empty-requirements.txt

You can adjust the rest of the template as needed. If you need a non-standard procedure, you can change the target of the template inputs to a local file. You must have a test_requirments file, as the manylinux wheel install test does not pull requirements when testing, and at least pytest is required.


To update, run:

git subtree pull --prefix .ci/azure-wheel-helpers master --squash

If you make changes inside the folder and want to contribute back, run:

git subtree push my_fixup_branch

As always, you can make a remote to shorten these commands.

Common needs

Using numpy with Cython

If you build with Cython, you will need to require an older version of Numpy. Either place this in your dev_requirements_file (classic builds) or your pyproject.toml (PEP 517 builds):

numpy==1.11.3; python_version<="3.5"
numpy==1.12.1; python_version=="3.6"
numpy==1.14.5; python_version=="3.7"
numpy==1.17.3; python_version>="3.8"

(Note: most of Scikit-HEP officially requires 1.13.3+, so you can simplify this with a single <='3.6')

Using PEP 517 builds

For PEP 517 builds, you need to have a pyproject.toml file. Then, for PIP > 10, the build happens in a custom environment that has only the packages you request. It replaces the deprecated and mostly non-functional setup_requires in, and even lets you select a build system other than setuptools. If you just use it as a replacement for setup_requires, you can still support pip < 10; users will just have to manually install the requirements (usually Numpy) beforehand. Here's an example of a Cython PEP 517 build:

requires = [
    "numpy==1.13.3; python_version<='3.6'",
    "numpy==1.14.5; python_version=='3.7'",
    "numpy==1.17.3; python_version>='3.8'",

Now, in, just import numpy and use it, no need to check to see if it there, etc.

Using Numpy parallel compile

If you have numpy available, you can add parallel compiles trivially:

# Use -j N or set the environment variable NPY_NUM_BUILD_JOBS
from numpy.distutils.ccompiler import CCompiler_compile
import distutils.ccompiler
distutils.ccompiler.CCompiler.compile = CCompiler_compile

Using Cython + Setuptools

Since setuptools>=18.0, you can now pass .pyx files directly as sources to Extension, and they get Cythonized for you! You just need Cython installed.


Copyright (c) 2019, Henry Schreiner.

Distributed under the 3-clause BSD license, see accompanying file LICENSE or for details.


No releases published


No packages published