Skip to content
Fetching contributors…
Cannot retrieve contributors at this time
585 lines (431 sloc) 20.2 KB

This file gives an overview of what is necessary to build binary releases for NumPy. Windows binaries are built here using Wine, they can of course also be built on Windows itself. Building OS X binaries on another platform is not possible, but our current OSX binary build procedure uses travis-ci virtual machines running OSX.

Current build and release info

The current info on building and releasing NumPy and SciPy is scattered in several places. It should be summarized in one place, updated and where necessary described in more detail. The sections below list all places where useful info can be found.

Source tree

  • INSTALL.txt

NumPy Docs wiki

Doc wiki

Release Scripts

Supported platforms and versions

Python 2.6-2.7 and >=3.2 are the currently supported versions when building from source. We test numpy against all these versions every time we merge code to trunk. Binary installers may be available for a subset of these versions (see below).


Python 2.7 and >=3.3 are the versions for which we provide binary installers. OS X versions >= 10.6 are supported. We build binary wheels for OSX that are compatible with Python, system Python, homebrew and macports - see this OSX wheel building summary for details.


32-bit Python 2.7, 3.3, 3.4 are the versions for which we provide binary installers. Windows XP, Vista and 7 are supported. Our current windows mingw toolchain is not able to build 64-bit binaries of numpy. We are hoping to update to a mingw-w64 toolchain soon.


Many distributions include NumPy. Building from source is also relatively straightforward. Only tarballs are created for Linux, no specific binary installers are provided (yet).

BSD / Solaris

No binaries are provided, but succesful builds on Solaris and BSD have been reported.

Tool chain


The same gcc version is used as the one with which Python itself is built on each platform. At the moment this means:

  • OS X builds on travis currently use clang. It appears that binary wheels for OSX >= 10.6 can be safely built from from OSX 10.9 when building against the Python from the installers.
  • Windows builds use MinGW 3.4.5. Updating this to a more recent MinGW with GCC 4.x is desired, but there are still practical difficulties in building the binary installers.

You will need Cython for building the binaries. Cython compiles the .pyx files in the numpy distribution to .c files.

Fortran: on OS X gfortran from this site is used. On Windows g77 (included in MinGW) is the current default, in the future this may shift to gfortran as well.


Building docs

  • Sphinx
  • numpydoc
  • Matplotlib
  • Texlive (or MikTeX on Windows)


For building Windows binaries on OS X Wine can be used. In Wine the following needs to be installed:

  • Python 2.6-2.7 and 3.3
  • MakeNsis
  • CpuId plugin for MakeNsis : this can be found in the NumPy source tree under tools/win32build/cpucaps and has to be built with MinGW (see SConstruct file in that dir for details)
  • MinGW
  • ATLAS, 3x ([No SSE, SSE2, SSE3] for superpack installer) : ATLAS does not compile under wine or on Windows out of the box. Binaries for ATLAS can be found in the vendor repository on GitHub (

To install Wine on OS X Snow Leopard the current options are to compile a current unstable version ,`<>`_, or to use an install script from here. For me, the former option did not work (everything compiled, but after installing Python the command import tempfile resulted in an exception. The latter option did work.

After successful installation and an invocation of the wine executable, a ~/.wine folder exists - new programs will be installed there in ~/.wine/drive_c. Installing Windows programs with .exe executables is done by running

$ wine yourprog.exe

and MSI installers can be installed with

$ msiexec /i yourprog.msi

For the above to work you probably need to put the wine-1.x.x/bin directory in your PATH.

To install MinGW, the easiest option is to use the automated installer on the MinGW download page. This will give you (at this moment) GCC 3.4.5; GCC 4.x is still not supported officially by MinGW.

To be able to use gcc and MakeNsis in Wine, the locations of gcc.exe and makensis.exe should be added to the Windows environment variable PATH. This can easily be done by running

$ wine regedit

add adding a PATH variable in HKEY_CURRENT_USER/Environment.


Virtualenv is a very useful tool to keep several versions of packages around. It is also used in the Paver script to build the docs.

What is released


Windows binary installers in "superpack" form for Python 2.7/3.3/3.4. A superpack contains three builds, for SSE2, SSE3 and no SSE.


OSX wheels built via travis-ci : see - see `building OSX wheels`_.


  • Release Notes
  • Changelog

Source distribution

We build source releases in both .zip and .tar.gz formats.

Release process

Agree on a release schedule

A typical release schedule is one beta, two release candidates and a final release. It's best to discuss the timing on the mailing list first, in order for people to get their commits in on time, get doc wiki edits merged, etc. After a date is set, create a new maintenance/x.y.z branch, add new empty release notes for the next version in the master branch and update the Trac Milestones.

Make sure current trunk builds a package correctly

git clean -fxd
python bdist
python sdist

To actually build the binaries after everything is set up correctly, the script can be used. For details of the build process itself it is best to read the script.


The following steps are repeated for the beta(s), release candidates(s) and the final release.

Check that docs can be built


cd doc/
make dist

to check that the documentation is in a buildable state. See doc/HOWTO_BUILD_DOCS.rst.txt for more details and for how to update

Check deprecations

Before the release branch is made, it should be checked that all deprecated code that should be removed is actually removed, and all new deprecations say in the docstring or deprecation warning at what version the code will be removed.

Check the C API version number

The C API version needs to be tracked in three places

  • numpy/core/
  • numpy/core/code_generators/cversions.txt
  • numpy/core/include/numpy/numpyconfig.h

There are three steps to the process.

  1. If the API has changed, increment the C_API_VERSION in The API is unchanged only if any code compiled against the current API will be backward compatible with the last released NumPy version. Any changes to C structures or additions to the public interface will make the new API not backward compatible.

  2. If the C_API_VERSION in the first step has changed, or if the hash of the API has changed, the cversions.txt file needs to be updated. To check the hash, run the script numpy/core/ and note the api hash that is printed. If that hash does not match the last hash in numpy/core/code_generators/cversions.txt the hash has changed. Using both the appropriate C_API_VERSION and hash, add a new entry to cversions.txt. If the API version was not changed, but the hash differs, you will need to comment out the previous entry for that API version. For instance, in NumPy 1.9 annotations were added, which changed the hash, but the API was the same as in 1.8. The hash serves as a check for API changes, but it is not definitive.

    If steps 1 and 2 are done correctly, compiling the release should not give a warning "API mismatch detect at the beginning of the build.

  3. The numpy/core/include/numpy/numpyconfig.h will need a new NPY_X_Y_API_VERSION macro, where X and Y are the major and minor version numbers of the release. The value given to that macro only needs to be increased from the previous version if some of the functions or macros in the include files were deprecated.

The C ABI version number in numpy/core/ should only be updated for a major release.

Check the release notes

Check that the release notes are up-to-date.

Write or update the release notes in a file named for the release, such as doc/release/1.11.0-notes.rst.

Mention at least the following:

  • major new features
  • deprecated and removed features
  • supported Python versions
  • for SciPy, supported NumPy version(s)
  • outlook for the near future

Also make sure that as soon as the branch is made, there is a new release notes file in trunk for the next release.

Update the release status and create a release "tag"

Identify the commit hash of the release, e.g. 1b2e1d63ff.

git co 1b2e1d63ff # gives warning about detached head

First, change/check the following variables in depending on the release version:

RELEASE_NOTES = 'doc/release/1.7.0-notes.rst'
LOG_START = 'v1.6.0'
LOG_END = 'maintenance/1.7.x'

Do any other changes. When you are ready to release, do the following changes:

diff --git a/ b/
index b1f53e3..8b36dbe 100755
--- a/
+++ b/
@@ -57,7 +57,7 @@ PLATFORMS           = ["Windows", "Linux", "Solaris", "Mac OS-
 MAJOR               = 1
 MINOR               = 7
 MICRO               = 0
-ISRELEASED          = False
+ISRELEASED          = True
 VERSION             = '%d.%d.%drc1' % (MAJOR, MINOR, MICRO)

 # Return the git revision as a string

And make sure the VERSION variable is set properly.

Now you can make the release commit and tag. We recommend you don't push the commit or tag immediately, just in case you need to do more cleanup. We prefer to defer the push of the tag until we're confident this is the exact form of the released code (see: :ref:`push-tag-and-commit`):

git commit -s -m "REL: Release." git tag -s <version>

The -s flag makes a PGP (usually GPG) signed tag. Please do sign the release tags.

The release tag should have the release number in the annotation (tag message). Unfortunately the name of a tag can be changed without breaking the signature, the contents of the message cannot.

See : for a discussion of signing release tags, and for instructions on creating a GPG key if you do not have one.

To make your key more readily identifiable as you, consider sending your key to public keyservers, with a command such as:

gpg --send-keys <yourkeyid>

Apply patch to fix bogus strides

NPY_RELAXED_STRIDE_CHECKING was made the default in Numpy 1.10 and bogus strides are used in the development branch to smoke out problems. The patch should be updated if necessary and applied to the release branch to rationalize the strides.

Update the version of the master branch

Increment the release number in Release candidates should have "rc1" (or "rc2", "rcN") appended to the X.Y.Z format.

Also create a new version hash in cversions.txt and a corresponding version define NPY_x_y_API_VERSION in numpyconfig.h

Trigger the OSX builds on travis

See build OSX wheels.

You may need to check the .travis.yml file of the repository.

Make sure that the releast tag has been pushed, and that the .travis.yml is set thusly:

- NP_COMMIT=latest-tag  # comment out to build version in submodule

Trigger a build by doing an empty (or otherwise) commit to the repository:

cd /path/to/numpy-wheels
git commit --allow-empty
git push

The wheels, once built, appear in

Trigger Windows builds on Appveyor

See: build Windows wheels

  • Clone / update the repository;
  • Check the appveyor.yml file in that repository;
  • Edit the line starting NP_VERSION: to give the numpy tag that you want to build;
  • Push up to github to trigger a build.

The wheels appear in a Rackspace CDN container at:

The contents via the HTTPS URL seems to get updated more slowly than via the HTTP URL, so if you need the binaries quickly, prefer the HTTP URL.

Trigger Manylinux builds on travis-ci


Until we move the manylinux build scripts, you'll need to ask @matthew-brett to make you a collaborator on the manylinux repos.

  • Clone / update the repository at
  • Edit the line in .travis.yml starting NUMPY_VERSIONS= to set the numpy tag to build;
  • Push your edits to .travis.yml up to github to trigger a mass manylinux build;
  • Clone / update the repository at;
  • Push an empty commit to the manylinux-testing repo to trigger a test run of the newly-built numpy wheels with a range of dependent libraries, as well as numpy's own unit tests. The tests will take several hours.

The built wheels will be available from a Rackspace CDN container at:

As for the other Rackspace containers, the HTTP address may update first, and you should wait 15 minutes after the build finishes before fetching the binaries. For the manylinux wheels, the time to run manylinux-testing is much greater than 15 minutes, so waiting for the tests to pass will be enough time for the binaries to refresh on Rackspace.

Make the release

Build the changelog and notes for upload with:

paver write_release_and_log

The tar-files and binary releases for distribution should be uploaded to SourceForge, together with the Release Notes and the Changelog. Uploading can be done through a web interface or, more efficiently, through scp/sftp/rsync as described in the SourceForge upload guide. For example:

scp <filename> <username>,<releasedir>/

Update PyPI

The wheels and source should be uploaded to PyPI.

You should upload the wheels first, and the source formats last, to make sure that pip users don't accidentally get a source install when they were expecting a binary wheel.

You can do this automatically using the wheel-uploader script from Here is the recommended incantation for downloading all the Windows, Manylinux, OSX wheels and uploading to PyPI.

cd ~/wheelhouse   # local directory to cache wheel downloads
wheel-uploader -u $MANYLINUX_URL -v -s -t manylinux1 numpy 1.11.1rc1
wheel-uploader -u $WINDOWS_URL -v -s -t win numpy 1.11.1rc1
wheel-uploader -u $OSX_URL -v -s -t macosx numpy 1.11.1rc1

The -v flag gives verbose feedback, -s causes the script to sign the wheels with your GPG key before upload.

You may well find that these uploads break at some point, with error messages from the PyPI server. In this case you'll have to continue the uploads by hand using twine, using something like:

twine upload -s numpy-1.11.1rc1-cp34-*.whl

Do this for the wheel files that wheel-uploader downloaded, but for which the upload failed.

The warehouse PyPI server seems to be more reliable in receiving automated wheel uploads. You can set the repository to upload to with the -r flag to wheel-uploader and twine. The warehouse repository URL for your ~/.pypirc file is

There are two ways to update the source release on PyPI, the first one is:

$ git clean -fxd  # to be safe
$ python sdist --formats=gztar,zip  # to check
# python sdist --formats=gztar,zip upload --sign

This will ask for your key PGP passphrase, in order to sign the built source packages.

The second way is to upload the PKG_INFO file inside the sdist dir in the web interface of PyPI. The source tarball can also be uploaded through this interface.

To push the travis-ci OSX wheels up to PyPI see :

Push the release tag and commit

Finally, now you are confident this tag correctly defines the source code that you released you can push the tag and release commit up to github:

git push  # Push release commit
git push upstream <version>  # Push tag named <version>

where upstream points to the main repository.


All documentation for a release can be updated on with:

make dist make upload USERNAME=<yourname> RELEASE=1.11.0

Note that <username> must have SSH credentials on the server. If you don't have those, ask someone who does (the list currently includes @rgommers, @juliantaylor and @pv).

Also rebuild and upload front page, if the release series is a new one. The front page sources have their own repo: Do the following:

  • Update index.rst for the new version.
  • make dist
  • Check that the built documentation is OK.
  • touch output-is-fine
  • make upload USERNAME=<username> RELEASE=1.x.y


A release announcement with a link to the download site should be placed in the sidebar of the front page of

The should be a PR at The file that needs modification is www/index.rst. Search for News.

Announce to the lists

The release should be announced on the mailing lists of NumPy and SciPy, to python-announce, and possibly also those of Matplotlib,IPython and/or Pygame.

During the beta/RC phase an explicit request for testing the binaries with several other libraries (SciPy/Matplotlib/Pygame) should be posted on the mailing list.

Announce to Linux Weekly News

Email the editor of LWN to let them know of the release. Directions at:

After the final release

After the final release is announced, a few administrative tasks are left to be done:

  • Forward port changes in the release branch to release notes and release scripts, if any, to trunk.
  • Update the Milestones in Trac.
Something went wrong with that request. Please try again.