Olivier Grisel
2016-04-03 20:26:51 UTC
Hi all,
Over the past couple of days Nathaniel Smith, Robert McGibbon, Matthew
Brett and I did a bunch of testing and bug fixing to get a working
build environment to generate binary packages for cython, numpy,
scipy, numpexpr, pandas, scikit-learn and statsmodels using an
embedded OpenBLAS 0.2.17 on Linux (only x86_64 and i686 platforms).
Here is Matthew's call to tests of the resulting packages on the numpy
mailing list:
https://mail.scipy.org/pipermail/numpy-discussion/2016-April/075234.html
You need pip 8.1 or later to install them (they are ignored by
previous versions of pip).
Please test those wheels on your machines, especially if you run
exotic Linux versions or less common CPU architectures.
If nobody has an objection I plan to upload the wheels for
scikit-learn 0.17.1 to PyPI as soon as numpy and scipy have their own
wheel packages uploaded there (probably during the coming week if no
blocker is reported in the mean time).
For those interested in the technical details, here are the tools that
we use to generate those wheels:
- manylinux is the official docker image that is used to build the
original wheels from source for
various versions of Python a controlled binary environment (based on
Centos 5.11 and gcc/g++/gfortran 4.8).
- auditwheel is a Python program to embed compiled dependencies into
the wheel (e.g. libopenblas.so)
https://github.com/pypa/auditwheel
- Matthew's script to build wheels for the last releases of the scipy
stack projects "en masse":
https://github.com/matthew-brett/manylinux-builds/
The manylinux1 platform tag it-self is documented in:
https://www.python.org/dev/peps/pep-0513/
Note that those wheels will not install on Alpine Linux or any other
Linux distribution that does not use glibc. Alpine Linux in particular
uses MUSL libc.
Happy testing!
Over the past couple of days Nathaniel Smith, Robert McGibbon, Matthew
Brett and I did a bunch of testing and bug fixing to get a working
build environment to generate binary packages for cython, numpy,
scipy, numpexpr, pandas, scikit-learn and statsmodels using an
embedded OpenBLAS 0.2.17 on Linux (only x86_64 and i686 platforms).
Here is Matthew's call to tests of the resulting packages on the numpy
mailing list:
https://mail.scipy.org/pipermail/numpy-discussion/2016-April/075234.html
You need pip 8.1 or later to install them (they are ignored by
previous versions of pip).
Please test those wheels on your machines, especially if you run
exotic Linux versions or less common CPU architectures.
If nobody has an objection I plan to upload the wheels for
scikit-learn 0.17.1 to PyPI as soon as numpy and scipy have their own
wheel packages uploaded there (probably during the coming week if no
blocker is reported in the mean time).
For those interested in the technical details, here are the tools that
we use to generate those wheels:
- manylinux is the official docker image that is used to build the
original wheels from source for
various versions of Python a controlled binary environment (based on
Centos 5.11 and gcc/g++/gfortran 4.8).
- auditwheel is a Python program to embed compiled dependencies into
the wheel (e.g. libopenblas.so)
https://github.com/pypa/auditwheel
- Matthew's script to build wheels for the last releases of the scipy
stack projects "en masse":
https://github.com/matthew-brett/manylinux-builds/
The manylinux1 platform tag it-self is documented in:
https://www.python.org/dev/peps/pep-0513/
Note that those wheels will not install on Alpine Linux or any other
Linux distribution that does not use glibc. Alpine Linux in particular
uses MUSL libc.
Happy testing!
--
Olivier
http://twitter.com/ogrisel - http://github.com/ogrisel
Olivier
http://twitter.com/ogrisel - http://github.com/ogrisel