The original Multibuild default was to upload wheels to a Rackspace container, where Rackspace kindly donated the hosting to the Scikit-learn team. We had a URL pointing to the Rackspace container: http://wheels.scipy.org.
Rackspace finally stopped subsidizing this container, and the Rackspace of http://wheels.scipy.org is no more. Some projects using Multibuild have moved to using https://anaconda.org/scipy-wheels-nightly/ for weekly uploads and https://anaconda.org/multibuild-wheels-staging for staging wheels to PyPI.
If you want to upload to Anaconda, and you don't need the extra storage space for nightly builds that Anaconda kindly donates to NumPy, SciPy etc, then you can do this with your own Anaconda organization.
See https://github.com/MacPython/nipy-wheels for a simple example.
Make an account at Anaconda.org.
Make an organization - for example we have used
nipy
.Navigate to
https://anaconda.org/nipy
(but use your organization).Go to your account menu towards the top left. This should be labeled with your organization name.
Select "Settings", then "Access".
Create an access token. Give it permission: "Allow write access to the API site", and (not sure if this is necessary) "Allow uploads to PyPI repositories".
"View" your token. It may be of form
ni-1234abcd-12ab-34dc-1234-d1e1f3a4b5c6
.Encrypt this to your repository, maybe using the
travis
command line tool (gem install travis
). Your command will be something like:travis encrypt -r MacPython/nipy-wheels ANACONDA_SECRET=ni-1234abcd-12ab-34dc-1234-d1e1f3a4b5c6
Note that
MacPython/nipy-wheels
is your Github organization/repository. The encryption only applies to Travis CI running against this repository.Go to your .travis.yml file and add the output
secure
key. This will look something like:env: global: # Following generated with # travis encrypt -r MacPython/nipy-wheels ANACONDA_SECRET=<the api key> # where <the api key> has API write access to the anaconda.org # organization named in $ANACONDA_ORG - secure: "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"
When Travis CI runs, this causes the
ANACONDA_SECRET
environment variable to contain the API key above.Also add this to your global environment variables:
- ANACONDA_ORG="nipy"
To upload from Travis CI, add a clause like this to the end of your
.travis.yml
:after_success: # Upload wheels to Anaconda.org - pip install git+https://github.com/Anaconda-Platform/anaconda-project - pip install git+https://github.com/Anaconda-Platform/anaconda-client - anaconda -t ${ANACONDA_SECRET} upload --force -u ${ANACONDA_ORG} ${TRAVIS_BUILD_DIR}/wheelhouse/*.whl
wheelhouse/*.whl
defines the files you want to upload.You might also want to build and upload from AppVeyor. To encrypt the API key above, go to https://ci.appveyor.com/account/matthew-brett/settings (where
matthew-brett
is the account from which your AppVeyor job runs. Click on "Encrypt YaML" on the left. Type in your API key value (e.g.ni-1234abcd-12ab-34dc-1234-d1e1f3a4b5c6
) as the value to encrypt. Click "Encrypt" and note the text it suggests. Now, at the top of yourappveyor.yml
file, add something like:environment: global: ANACONDA_ORG: "nipy" ANACONDA_SECRET: secure: Ds0PkQD0b/QOfoNoiPuFJb01zg0Mq0dkAxIG2jRXocCAereSXdWw6XYaDrutHWme
where
secure:
is the text suggested from "Encrypt" above, andnipy
is your Anaconda organization.Finally, add a clause like the following to the end of your
appveyor.yml
file:on_success: # Upload the generated wheel package to Anaconda.org - pip install git+https://github.com/Anaconda-Platform/anaconda-project - pip install git+https://github.com/Anaconda-Platform/anaconda-client - anaconda -t %ANACONDA_SECRET% upload --force -u %ANACONDA_ORG% nipy\dist\*.whl
where
nipy\dist\*.whl
finds the files you want to upload.
There's a simple example of these steps applied at https://github.com/MacPython/nipy-wheels.
Here is the NumPy code (for running on Travis CI) to upload to Anaconda: https://github.com/MacPython/numpy-wheels/blob/master/.travis.yml#L99
For projects housed under the MacPython GitHub organization, you have access to Anaconda upload tokens via the "Organization Secrets" https://github.com/MacPython/numexpr-wheels/settings/secrets . You can use these to move to GitHub Actions (they provide x86 machines for Windows, Linux and Mac). Otherwise we (please raise an issue here) will need to negotiate getting you tokens, or you can make your own, as above.
Another option is to use GitHub for staging --- as do Cython for Travis CI and for AppVeyor.
A set of scripts to automate builds of macOS and Manylinux1 wheels on the Travis CI infrastructure, and also Windows wheels on the AppVeyor infrastructure.
The Travis CI scripts are designed to build and test:
- 64-bit macOS wheels built for macOS 10.9+
- 64/32-bit macOS wheels built for macOS 10.6+
- 64-bit
manylinuxX_x86_64
wheels, both narrow and wide Unicode builds, where X is any valid Manylinux version: 1, 2010, 2014, _2_24 or _2_28. - 32-bit
manylinuxX_i686
wheels, both narrow and wide Unicode builds
You can currently build and test against Pythons 2.7, 3.5, 3.6, 3.7, 3.8, 3.9, 3.10 and 3.11.
The small innovation here is that you can test against Linux 32-bit builds, both wide and narrow Unicode Python 2 builds, which was not easy on the default Travis CI configurations.
The AppVeyor setup is designed to build and test:
- 64-bit Windows
win_amd64
wheels - 32-bit Windows
win32
wheels
You can currently build and test against Pythons 2.7, 3.5, 3.6, 3.7, 3.8, 3.9, 3.10 and 3.11.
Multibuild is a series of bash scripts that define bash functions to build and test wheels.
Configuration is by overriding the default build function, and defining a test function.
The bash scripts are layered, in the sense that they are composed of a number of scripts which are sourced in sequence, each one potentially overriding previous ones.
The following bash scripts are sourced in this order:
multibuild/common_utils.sh multibuild/osx_utils.sh env_vars.sh multibuild/configure_build.sh multibuild/library_builders.sh config.sh
See multibuild/travis_osx_steps.sh
The macOS build / test phases run on the macOS VM started by Travis CI.
Therefore any environment variable defined in .travis.yml
or the bash
shell scripts listed above are available for your build and test.
Build options are controlled mainly by the following environment variables:
MB_PYTHON_VERSION
sets the Python version targeted:major.minor.patch
for CPython, orpypy-major.minor
for PyPy.MB_PYTHON_OSX_VER
sets the minimum macOS SDK version for any C extensions. For CPython targets it may be set to 10.6, 10.9 or 11.0, provided a corresponding Python build is available at python.org. It defaults to the highest version available. It's ignored for PyPy targets.PLAT
sets the architectures built for any C extensions:x86_64
orintel
for 64-bit or 64/32-bit respectively. It defaults to the same arches as the target Python version: arm64 for macOS 11.0; x86_64 for CPython macOS 10.9 or PyPy; and 64/32-bit for CPython 10.6.
In most cases it's best to rely on the defaults for MB_PYTHON_OSX_VER
and
PLAT
, rather than setting them explicitly. Examples of exceptions to this
guideline include:
- setting
MB_PYTHON_OSX_VER=10.6
to build a 10.6 64/32-bit CPython wheel for Python 2.7 (default for 2.7 is 10.9 64-bit) - setting
MB_PYTHON_OSX_VER=10.6 and PLAT=x86_64
to build a 10.6 64-bit only wheel (10.6 would normally be 64/32-bit). Such a wheel would still have a platform tag ofmacosx_10_6_intel
, advertising support for both 64 and 32-bit, but wouldn't work in 32-bit mode. This may be OK given how unlikely it is that there is still anyone actually running Python on macOS in 32-bit mode.
The build_wheel
function builds the wheel, and install_run
function installs and tests it. Look in multibuild/common_utils.sh
for
default definitions of these functions. See below for more details, many of
which are common to macOS and Linux.
The build phase is in a Manylinux Docker container, but the test phase is in a clean container.
Specify the Manylinux version to build for with the MB_ML_VER
environment
variable. The default version is 2014
. Versions that are currently valid are:
1
corresponding to manylinux1 (see PEP 513).2010
corresponding to manylinux2010 (see PEP 571).2014
corresponding to manylinux2014 and adds more architectures toPLAT
(see PEP 599)._2_24
corresponding to manylinux_2_24 (see PEP 600)._2_28
corresponding to manylinux_2_28 (see PEP 600).
The environment variable specified which Manylinux docker container you are building in.
The PLAT
environment variable can be one of
x86_64
, for 64-bit x86i686
, for 32-bit x86s390x
, for 64-bit s390xppc64le
, for PowerPCaarch64
, for ARMarm64
, for Apple siliconuniversal2
, for both Apple silicon and 64-bit x86
The default is x86_64
. Only x86_64
and i686
are valid on manylinux1 and manylinux2010.
multibuild/travis_linux_steps.sh
defines the build_wheel
function,
which starts up the Manylinux1 Docker container to run a wrapper script
multibuild/docker_build_wrap.sh
, that (within the container) sources the
following bash scripts:
multibuild/common_utils.sh multibuild/manylinux_utils.sh env_vars.sh multibuild/configure_build.sh multibuild/library_builders.sh config.sh
See docker_build_wrap.sh
to review the order of script sourcing.
See the definition of build_multilinux
in
multibuild/travis_linux_steps.sh
for the environment variables passed from
Travis CI to the Manylinux1 container.
Once in the container, after sourcing the scripts above, the wrapper runs the
real build_wheel
function, which now comes (by default) from
multibuild/common_utils.sh
.
Specify the version to test with the DOCKER_TEST_IMAGE
environment
variable. The default version is dependent on MB_ML_LIBC
and PLAT
.
When MB_ML_LIBC
is musllinux
:
multibuild/alpine3.20_x86_64
, whenPLAT
isx86_64
multibuild/alpine3.20_arm64v8
, whenPLAT
isaarch64
Otherwise:
multibuild/focal_x86_64
, whenPLAT
isx86_64
matthewbrett/trusty:32
whenPLAT
isi686
(Yes, an older image for 32-bit)multibuild/focal_arm64v8
whenPLAT
isaarch64
multibuild/focal_ppc64le
whenPLAT
isppc64le
multibuild/focal_s390x
whenPLAT
iss390x
Other valid values are any in https://hub.docker.com/orgs/multibuild/repositories,
using the correct platform code. Alternatively, you can use the substitution
pattern multibuild/focal_{PLAT}
in the .travis.yml
file.
See multibuild/docker_test_wrap.sh
.
multibuild/travis_linux_steps.sh
defines the install_run
function,
which starts up the testing Docker container with the wrapper script
multibuild/docker_test_wrap.sh
. The wrapper script sources the following
bash scripts:
multibuild/common_utils.sh config.sh
See docker_test_wrap.sh
for script source order.
See install_run
in multibuild/travis_linux_steps.sh
for the
environment variables passed into the container.
It then (in the container) runs the real install_run
command, which comes
(by default) from multibuild/common_utils.sh
.
The standard build command is build_wheel
. This is a bash function. By
default the function that is run on macOS, and in the Manylinux container for
the build phase, is defined in multibuild/common_utils.sh
. You can
override the default function in the project config.sh
file (see below).
If you are building a wheel from PyPI, rather than from a source repository,
you can use the build_index_wheel
command, again defined in
multibuild/common_utils.sh
.
Typically, you can get away with leaving the default build_wheel
/
build_index_wheel
functions to do their thing, but you may need to define
a pre_build
function in config.sh
. The default build_wheel
and
build_index_wheel
functions will call the pre_build
function, if
defined, before building the wheel, so pre_build
is a good place to build
any required libraries.
The standard test command is the bash function install_run
. The version
run on macOS and in the Linux testing container is also defined in
multibuild/common_utils.sh
. Typically, you do not override this function,
but you in that case you will need to define a run_tests
function, to run
your tests, returning a non-zero error code for failure. The default
install_run
implementation calls the run_tests
function, which you
will likely define in config.sh
. See the examples below for examples of
less and more complicated builds, where the complicated builds override more
of the default implementations.
Make a repository for building wheels on Travis CI - e.g. https://github.com/MacPython/astropy-wheels - or in your case maybe
https://github.com/your-org/your-project-wheels
;Add this (here) repository as a submodule:
git submodule add https://github.com/multi-build/multibuild.git
Add your own project repository as another submodule:
git submodule add https://github.com/your-org/your-project.git
Create a
.travis.yml
file, something like this:env: global: - REPO_DIR=your-project # Commit from your-project that you want to build - BUILD_COMMIT=v0.1.0 # pip dependencies to _build_ your project - BUILD_DEPENDS="cython numpy" # pip dependencies to _test_ your project. Include any dependencies # that you need, that are also specified in BUILD_DEPENDS, this will be # a separate install. # Now see the Uploads section for the stuff you need to # upload your wheels after CI has built them. # You will likely prefer "language: generic" for travis configuration, # rather than, say "language: python". Multibuild doesn't use # Travis-provided Python but rather installs and uses its own, where the # Python version is set from the MB_PYTHON_VERSION variable. You can still # specify a language here if you need it for some unrelated logic and you # can't use Multibuild-provided Python or other software present on a # builder. language: generic # For CPython macOS builds only, the minimum supported macOS version and # architectures of any C extensions in the wheel are set with the variable # MB_PYTHON_OSX_VER: 10.9 (64-bit only) or 10.6 (64/32-bit dual arch). By # default this is set to the highest available for the Python version selected # using MB_PYTHON_VERSION. You should only need to set this explicitly if you # are building a 10.6 dual-arch build for a CPython version where both a 10.9 and # 10.6 build are available (for example, 2.7 or 3.7). # All PyPy macOS builds are 64-bit only. # Required in Linux to invoke `docker` ourselves services: docker # Host distribution. This is the distribution from which we run the build # and test containers, via docker. dist: xenial # osx image that enables building Apple silicon libraries osx_image: xcode12.2 matrix: include: - os: linux env: MB_PYTHON_VERSION=2.7 - os: linux env: - MB_PYTHON_VERSION=2.7 - UNICODE_WIDTH=16 - os: linux env: - MB_PYTHON_VERSION=2.7 - PLAT=i686 - os: linux env: - MB_PYTHON_VERSION=2.7 - PLAT=i686 - UNICODE_WIDTH=16 - os: linux env: - MB_PYTHON_VERSION=3.5 - os: linux env: - MB_PYTHON_VERSION=3.5 - PLAT=i686 - os: linux env: - MB_PYTHON_VERSION=3.6 - os: linux env: - MB_PYTHON_VERSION=3.6 - PLAT=i686 - os: osx env: - MB_PYTHON_VERSION=2.7 - MB_PYTHON_OSX_VER=10.6 - os: osx env: - MB_PYTHON_VERSION=2.7 - os: osx env: - MB_PYTHON_VERSION=3.5 - os: osx env: - MB_PYTHON_VERSION=3.6 - os: osx env: - MB_PYTHON_VERSION=3.7 - MB_PYTHON_OSX_VER=10.6 - os: osx env: - MB_PYTHON_VERSION=3.7 - os: osx env: - MB_PYTHON_VERSION=3.8 - os: osx env: - MB_PYTHON_VERSION=3.9 - PLAT="universal2" - os: osx env: - MB_PYTHON_VERSION=3.9 - os: osx language: generic env: - MB_PYTHON_VERSION=pypy-5.7 before_install: - source multibuild/common_utils.sh - source multibuild/travis_steps.sh - before_install install: # Maybe get and clean and patch source - clean_code $REPO_DIR $BUILD_COMMIT - build_wheel $REPO_DIR $PLAT script: - install_run $PLAT after_success: # Here you should put the code to upload your wheels # See the Uploads section for more details.
The example above is for a project building from a Git submodule. If you aren't building from a submodule, but want to use
pip
to build from a source archive on https://pypi.org or similar, replace the first few lines of the.travis.yml
file with something like:env: global: # Instead of REPO_DIR, BUILD_COMMIT - PROJECT_SPEC="tornado==4.1.1"
then your
install
section could look something like this:install: - build_index_wheel $PROJECT_SPEC
Next create a
config.sh
for your project, that fills in any steps you need to do before building the wheel (such as building required libraries). You also need this file to specify how to run your tests:# Define custom utilities # Test for macOS with [ -n "$IS_MACOS" ] function pre_build { # Any stuff that you need to do before you start building the wheels # Runs in the root directory of this repository. : } function run_tests { # Runs tests on installed distribution from an empty directory python --version python -c 'import sys; import yourpackage; sys.exit(yourpackage.test())' }
Optionally you can specify a different location for
config.sh
file with the$CONFIG_PATH
environment variable.Optionally, create an
env_vars.sh
file to override the defaults for any environment variables used byconfigure_build.sh
/library_builders.sh
. In Linux, the environment variables used for the build cannot be set in the.travis.yml
file, because the build processing runs in a Docker container, so the only environment variables that reach the container are those passed in via thedocker run
command, or those set inenv_vars.sh
.As for the
config.sh
file, you can specify a different location for the file by setting the$ENV_VARS_PATH
environment variable. The path in$ENV_VARS_PATH
is relative to the repository root directory. For example, if your repository had a subdirectoryscripts
with a filemy_env_vars.sh
, you should setENV_VARS_PATH=scripts/my_env_vars.sh
.Make sure your project is set up to build on Travis CI, and you should now be ready (to begin the long slow debugging process, probably).
For the Windows wheels, create an
appveyor.yml
file, something like:- https://github.com/MacPython/astropy-wheels/blob/master/appveyor.yml
- https://github.com/MacPython/nipy-wheels/blob/master/appveyor.yml
- https://github.com/MacPython/pytables-wheels/blob/master/appveyor.yml
Note the Windows test customizations etc are inside
appveyor.yml
, and thatconfig.sh
andenv_vars.sh
are only for the Linux/Mac builds on Travis CI.Make sure your project is set up to build on AppVeyor, and you should now be ready (for what could be another round of slow debugging).
For Apple silicon support you can either create an
arm64
wheel or auniversal2
wheel by supplyingPLAT
env variable.universal2
builds work on botharm64
andx86_64
platforms and also make it possible for the wheel code to work when switching the architecture on Apple silicon machines wherex86_64
can be run using Rosetta2 emulation.There are two ways to build
universal2
builds.- Build with
-arch x86_64 -arch arm64
. These flags instruct the C/C++ compiler to compile twice and create a fat object/executable/library. This is the easiest, but has several drawbacks. If you are using C/C++ libraries that are built using library_builders, it's highly likely that they don't build correctly because most build systems and packages don't support building fat binaries. We could possibly build them separately and fuse them, but the headers might not be identical which is required when building the wheel as auniversal2
wheel. If you are using Fortran,gfortran
doesn't support fat binaries. - Build
arm64
andx86_64
wheels separately and fuse them. For this to work, we need to build the C/C++ libraries twice. Therefore, the library building is once called withBUILD_PREFIX=${BUILD_PREFIX:-/usr/local}
forx86_64
and then called again withBUILD_PREFIX=/opt/arm64-builds
. Once the two wheels are created, these two are merged. Both thearm64
anduniversal2
wheels are outputs for this build.
In multibuild we are going with option 2. You can override this behaviour by overriding the function
wrap_wheel_builder
. To build Apple silicon builds, you should use a CI service with Xcode 12 with universal build support and make sure that xcode is the default.- Build with
If your project depends on NumPy, you will want to build against the earliest NumPy that your project supports - see forward, backward NumPy compatibility. See the astropy-wheels Travis file for an example specifying NumPy build and test dependencies.
Here are some simple example projects:
- https://github.com/MacPython/astropy-wheels
- https://github.com/scikit-image/scikit-image-wheels
- https://github.com/MacPython/nipy-wheels
- https://github.com/MacPython/dipy-wheels
Less simple projects where there are some serious build dependencies, and / or macOS / Linux differences:
- https://github.com/MacPython/matplotlib-wheels
- https://github.com/python-pillow/Pillow-wheels
- https://github.com/MacPython/h5py-wheels
The main multibuild repository is always at https://github.com/multi-build/multibuild
We try to keep the master
branch stable and do testing and development
in the devel
branch. From time to time we merge devel
into master
.
In practice, you can check out the newest commit from devel
that works
for you, then stay at it until you need newer features.