Knowledge Base

Using CMake

CMake can be used to build more complex projects in or bld.bat scripts.

If you are using cmake, be sure to make it a build requirement in the build section. You may also need to include make or ninja depending on your platform and build tools. On Windows, you can also use nmake to build, but that does not need to be explicitly included.

    - cmake
    - make  # [not win]
    - ninja  # [win]

For CMake projects using the FindPython module, you can tell CMake which Python to use by passing -DPython_EXECUTABLE="$PYTHON" (macOS or Linux) or -DPython_EXECUTABLE="%PYTHON%" (Windows) as a command line option. Older CMake projects may require similar, but slightly different options.


Don’t forget that depending on which CMake module you use you have to use a different command:

or if you are still on the deprecated FindPythonLibs: -DPYTHON_EXECUTABLE=....

Some optional, but useful CMake options:

  • -DCMAKE_BUILD_TYPE=Release Configure as release build. This is better done on the initial cmake call as some packages construct different build configurations depending on this flag.

  • -DCMAKE_INSTALL_PREFIX=$PREFIX Specify the install location.

  • -DCMAKE_INSTALL_LIBDIR=lib Libraries will land in $PREFIX/lib, sometimes projects install into lib64 or similar but on conda-forge we keep shared libraries in simply lib.

  • -DBUILD_SHARED_LIBS=ON Instruct CMake to build shared libraries instead of static ones.

  • ${CMAKE_ARGS} Add variables defined by conda-forge internally. This is required to enable various conda-forge enhancements, like CUDA builds.

Here are some basic commands for you to get started. These are dependent on your source code layout and aren’t intended to be used “as is”.

CMake lines for (macOS/Linux):

cmake CMakeLists.txt -DPython3_EXECUTABLE="$PYTHON"
cmake --build . --config Release

CMake lines for bld.bat (Windows):

cmake -G "NMake Makefiles" -DCMAKE_BUILD_TYPE=Release -DPython3_EXECUTABLE="%PYTHON%"
if errorlevel 1 exit /b 1
cmake --build . --config Release
if errorlevel 1 exit /b 1

See also the bld.bat in the Windows section below for an additional example.

Other useful cmake options are -B<directory> and -S<directory> to specify build and source directories.

Moving from an autotools build to a CMake build

Some packages maintain an autotools build and a cmake build. Some maintainers would like to switch to a cmake build because that provides windows builds easily. These builds are mostly not ABI compatible with each other. Here are some things you should check,

  1. Check that both libraries have the same SONAME on linux

    Run readelf -d /path/to/

  2. Check that both libraries have the same install name and have the same compatibility and current versions.

    Run otool -L /path/to/lib.dylib. The second line should give you the three pieces of information

  3. Check that the file list is the same in both.

  4. Check that you use the same options as the same autoconf build.

  5. Check that the symbols exported are the same.

  6. Check that additional packaging information stays the same, e.g. is the same pkg-config information provided.

Particularities on Windows

This document presents conda-forge and conda-build information and examples while building on Windows.

Local testing

The first thing that you should know is that you can locally test Windows builds of your packages even if you don’t own a Windows machine. Microsoft makes available free, official Windows virtual machines (VMs) at this website. If you are unfamiliar with VM systems or have trouble installing Microsoft’s VMs, please use a general web search to explore — while these topics are beyond the scope of this documentation, there are ample discussions on them on the broader Internet.

In order to compile native code (C, C++, etc.) on Windows, you will need to install Microsoft’s Visual C++ build tools on your VM. You must install particular versions of these tools — this is to maintain compatibility between compiled libraries used in Python, as described on this Python wiki page. The current relevant versions are:

  • For Python 3.5–3.7: Visual C++ 14.0

While you can obtain these tools by installing the right version of the full Visual Studio development environment, you can save a lot of time and bandwidth by installing standalone “build tools” packages. The links are as follows:

If you need more information. Please refer the Python wiki page on Windows compilers.

Simple CMake-Based bld.bat

Some projects provide hooks for CMake to build the project. The following example bld.bat file demonstrates how to build a traditional, out-of-core build for such projects.

CMake-based bld.bat:

setlocal EnableDelayedExpansion

:: Make a build folder and change to it.
mkdir build
cd build

:: Configure using the CMakeFiles
cmake -G "NMake Makefiles" ^
if errorlevel 1 exit 1

:: Build!
if errorlevel 1 exit 1

:: Install!
nmake install
if errorlevel 1 exit 1

The following feedstocks are examples of this build structure deployed:

Building for different VC versions

On Windows, different Visual C versions have different ABI and therefore a package needs to be built for different Visual C versions. Packages are tied to the VC version that they were built with and some packages have specific requirements of the VC version. For example, python 2.7 requires vc 9 and python 3.5 requires vc 14.

With conda-build 3.x, vc can be used as a selector when using the compiler jinja syntax.

    - {{ compiler('cxx') }}

To skip building with a particular vc version, add a skip statement.

    skip: true  # [win and vc<14]

    - {{ compiler('cxx') }}

Using vs2019

To use vs2019 make the following changes:

In conda_build_config.yaml file:

- vs2019
- vs2019

In conda-forge.yml file:

          vmImage: windows-2019

For example see the changes made in the conda_build_config.yaml and conda-forge.yml files in this PR.

After making these changes don’t forget to rerender with conda-smithy (to rerender manually use conda smithy rerender from the command line).

Special Dependencies and Packages


Compilers are dependencies with a special syntax and are always added to requirements/build.

There are currently three supported compilers:

  • C

  • cxx

  • Fortran

A package that needs all three compilers would define

    - {{ compiler('c') }}
    - {{ compiler('cxx') }}
    - {{ compiler('fortran') }}


Appropriate compiler runtime packages will be automatically added to the package’s runtime requirements and therefore there’s no need to specify libgcc or libgfortran. There are additional informations about how conda-build 3 treats compilers in the conda docs.

Core Dependency Tree Packages (CDTs)

Dependencies outside of the conda-forge channel should be avoided (see Avoid external dependencies). However, there are a few exceptions:

Some dependencies are so close to the system that they are not packaged with conda-forge. These dependencies have to be satisfied with Core Dependency Tree (CDT) packages.

A CDT package consists of repackaged CentOS binaries from the appropriate version, either 6 or 7 depending on user choice and platform. We manage the build of CDT packages using a centralized repo, conda-forge/cdt-builds, as opposed to generating feedstocks for them. (Note that historically we did use feedstocks but this practice has been deprecated.) To add a new CDT, make a PR on the conda-forge/cdt-builds repo.

In conda-forge the primary usages of CDTs is currently for packages that link against libGL.


In addition to the required compilers {{ compiler('c') }} and/or {{ compiler('cxx') }}, the following CDT packages are required for linking against libGL:

    - {{ cdt('mesa-libgl-devel') }}  # [linux]
    - {{ cdt('mesa-dri-drivers') }}  # [linux]
    - {{ cdt('libselinux') }}  # [linux]
    - {{ cdt('libxdamage') }}  # [linux]
    - {{ cdt('libxxf86vm') }}  # [linux]
    - {{ cdt('libxext') }}     # [linux]
    - xorg-libxfixes  # [linux]

If you need a fully functional binary in the test phase, you have to also provide the shared libraries via yum_requirements.txt (see yum_requirements.txt).


You will need to re-render the feedstock after making these changes.

Building Against NumPy

Packages that link against NumPy need special treatment in the dependency section. Finding numpy.get_include() in or cimport statements in .pyx or .pyd files are a telltale sign that the package links against NumPy.

In the case of linking, you need to use the pin_compatible function to ensure having a compatible numpy version at run time:

  - numpy
  - {{ pin_compatible('numpy') }}

At the time of writing, above is equivalent to the following,

  - numpy 1.14.6
  - numpy >=1.14.6,<2.0.a0


1. You still need to respect minimum supported version of numpy for the package! That means you cannot use numpy 1.9 if the project requires at least numpy 1.12, adjust the minimum version accordingly!

  - numpy 1.12.*
  - {{ pin_compatible('numpy') }}

2. if your package supports numpy 1.7, and you are brave enough :-), there are numpy packages for 1.7 available for Python 2.7 in the channel.

JupyterLab Extensions

A typical JupyterLab extension has both Python and JavaScript components. These should be packaged together, to prevent node from being needing to grab the JavaScript side of the package on the user’s machine. To package an extension, the build should have the following meta.yaml snippet:

  noarch: python

    - python
    - nodejs
    - pip
    - python
    - nodejs
    - jupyterlab >=2

Please use the following script in your recipe:

#!/usr/bin/env bash
set -ex

$PYTHON -m pip install . -vv
npm pack ${PKG_NAME}@${PKG_VERSION}
mkdir -p ${PREFIX}/share/jupyter/lab/extensions/js
cp ${PKG_NAME}-${PKG_VERSION}.tgz ${PREFIX}/share/jupyter/lab/extensions/js

Since this is a noarch recipe, the build script only needs to run on linux-64. Also note that we do not need to run jupyter labextension install or jupyter lab build as part of the package build or in any post-link scripts. This is because JupyterLab will run the build step itself when it is next run. The ${PREFIX}/share/jupyter/lab/extensions/js directory which JupyterLab knows to build from when performing this build step.

Message passing interface (MPI)


This section originates from Min’s notes:

MPI Variants in conda-forge

How are MPI variants best handled in conda-forge?

There are a few broad cases:

  • package requires a specific MPI provider (easy!)

  • the package works with any MPI provider (e.g. mpich, openmpi)

  • the package works with/without MPI

Note that sometimes users want to use packages in conda-forge built against our MPI libraries but linked to external MPI libraries at runtime. If you are interested in this procedure, see Using External Message Passing Interface (MPI) Libraries for details.

Building MPI variants

In conda_build_config.yaml:

  - mpich
  - openmpi

In meta.yaml:

    - {{ mpi }}

And rerender with:

conda-smithy rerender -c auto

to produce the build matrices.

Including a no-mpi build

Some packages (e.g. hdf5) may want a no-mpi build, in addition to the mpi builds. To do this, add nompi to the mpi matrix:

  - nompi
  - mpich
  - openmpi

and apply the appropriate conditionals in your build:

    - {{ mpi }}  # [mpi != 'nompi']
    - {{ mpi }}  # [mpi != 'nompi']
Preferring a provider (usually nompi)

Up to here, mpi providers have no explicit preference. When choosing an MPI provider, the mutual exclusivity of the mpi metapackage allows picking between mpi providers by installing an mpi provider, e.g.

conda install mpich ptscotch


conda install openmpi ptscotch

This doesn’t extend to nompi, because there is no nompi variant of the mpi metapackage. And there probably shouldn’t be, because some packages built with mpi don’t preclude other packages in the env that may have an mpi variant from using the no-mpi variant of the library (e.g. for a long time, fenics used mpi with no-mpi hdf5 since there was no parallel hdf5 yet. This works fine, though some features may not be available).

Typically, if there is a preference it will be for the serial build, such that installers/requirers of the package only get the mpi build if explicitly requested. We use a higher build number for the nompi variant in this case.

Here is an example build section:

{% if mpi == 'nompi' %}
# prioritize nompi variant via build number
{% set build = build + 100 %}
{% endif %}
  number: {{ build }}

  # add build string so packages can depend on
  # mpi or nompi variants explicitly:
  # `pkg * mpi_mpich_*` for mpich
  # `pkg * mpi_*` for any mpi
  # `pkg * nompi_*` for no mpi

  {% if mpi != 'nompi' %}
  {% set mpi_prefix = "mpi_" + mpi %}
  {% else %}
  {% set mpi_prefix = "nompi" %}
  {% endif %}
  string: "{{ mpi_prefix }}_h{{ PKG_HASH }}_{{ build }}"


{{ PKG_HASH }} avoids build string collisions on most variants, but not on packages that are excluded from the default build string, e.g. Python itself. If the package is built for multiple Python versions, use:

string: "{{ mpi_prefix }}_py{{ py }}h{{ PKG_HASH }}_{{ build }}"

as seen in mpi4py

This build section creates the following packages:

  • pkg-x.y.z-mpi_mpich_h12345_0

  • pkg-x.y.z-mpi_openmpi_h23456_0

  • pkg-x.y.z-nompi_h34567_100

Which has the following consequences:

  • The nompi variant is preferred, and will be installed by default unless an mpi variant is explicitly requested.

  • mpi variants can be explicitly requested with pkg=*=mpi_{{ mpi }}_*

  • any mpi variant, ignoring provider, can be requested with pkg=*=mpi_*

  • nompi variant can be explicitly requested with pkg=*=nompi_*

If building with this library creates a runtime dependency on the variant, the build string pinning can be added to run_exports.

For example, if building against the nompi variant will work with any installed version, but building with a given mpi provider requires running with that mpi:

  {% if mpi != 'nompi' %}
    - {{ name }} * {{ mpi_prefix }}_*
  {% endif %}

Remove the if mpi... condition if all variants should create a strict runtime dependency based on the variant chosen at build time (i.e. if the nompi build cannot be run against the mpich build).

Complete example

Combining all of the above, here is a complete recipe, with:

  • nompi, mpich, openmpi variants

  • run-exports to apply mpi choice made at build time to runtime where nompi builds can be run with mpi, but not vice versa.

  • nompi variant is preferred by default

  • only build nompi on Windows

This matches what is done in hdf5.

# conda_build_config.yaml
  - nompi
  - mpich  # [not win]
  - openmpi  # [not win]
# meta.yaml
{% set name = 'pkg' %}
{% set build = 0 %}

# ensure mpi is defined (needed for conda-smithy recipe-lint)
{% set mpi = mpi or 'nompi' %}

{% if mpi == 'nompi' %}
# prioritize nompi variant via build number
{% set build = build + 100 %}
{% endif %}

  number: {{ build }}

  # add build string so packages can depend on
  # mpi or nompi variants explicitly:
  # `pkg * mpi_mpich_*` for mpich
  # `pkg * mpi_*` for any mpi
  # `pkg * nompi_*` for no mpi

  {% if mpi != 'nompi' %}
  {% set mpi_prefix = "mpi_" + mpi %}
  {% else %}
  {% set mpi_prefix = "nompi" %}
  {% endif %}
  string: "{{ mpi_prefix }}_h{{ PKG_HASH }}_{{ build }}"

  {% if mpi != 'nompi' %}
    - {{ name }} * {{ mpi_prefix }}_*
  {% endif %}

    - {{ mpi }}  # [mpi != 'nompi']
    - {{ mpi }}  # [mpi != 'nompi']

And then a package that depends on this one can explicitly pick the appropriate mpi builds:

# meta.yaml

    - {{ mpi }}  # [mpi != 'nompi']
    - pkg
    - pkg * mpi_{{ mpi }}_*  # [mpi != 'nompi']
    - {{ mpi }}  # [mpi != 'nompi']
    - pkg * mpi_{{ mpi }}_*  # [mpi != 'nompi']

mpi-metapackage exclusivity allows mpi_* to resolve the same as mpi_{{ mpi }}_* if {{ mpi }} is also a direct dependency, though it’s probably nicer to be explicit.

Just mpi example

Without a preferred nompi variant, recipes that require mpi are much simpler. This is all that is needed:

# conda_build_config.yaml
  - mpich
  - openmpi
# meta.yaml
    - {{ mpi }}
    - {{ mpi }}


You can enable OpenMP on macOS by adding the llvm-openmp package to the build section of the meta.yaml. For Linux OpenMP support is on by default, however it’s better to explicitly depend on the libgomp package which is the OpenMP implementation from the GNU project.

# meta.yaml
    - llvm-openmp  # [osx]
    - libgomp      # [linux]

Switching OpenMP implementation

On macOS, only LLVM’s OpenMP implementation llvm-openmp is supported. This implementation is used even in Fortran code compiled using GNU’s gfortran.

On Linux (except aarch64), packages are linked against GNU’s, but the OpenMP library at install time can be switched from GNU to LLVM by doing the following.

conda install _openmp_mutex=*=*_llvm

OpenMP library can be switched back to GNU’s libgomp by doing the following.

conda install _openmp_mutex=*=*_gnu


OpenMP library switching is possible because LLVM’s implementation has the symbol’s from GNU in addition to the LLVM ones (originally from Intel). An object file generated by gcc, g++ or gfortran will have GNU’s symbols and therefore the underlying library can be switched. However, an object file generated by clang or clang++ will have LLVM’s symbols and therefore the underlying OpenMP library cannot be switched to GNU’s library.

One reason you may wish to switch to LLVM is because the implementation is fork safe. One reason to keep using the GNU implementation is that the OpenMP target offloading symbols in libgomp like GOMP_target are empty stubs in LLVM and therefore does not work.


Dependencies can be installed into the build container with yum, by listing package names line by line in a file named yum_requirements.txt in the recipe directory of a feedstock.

There are only very few situations where dependencies installed by yum are acceptable. These cases include

  • satisfying the requirements of CDT packages during test phase

  • installing packages that are only required for testing

After changing yum_requirements.txt, rerender to update the configuration.


If a package needs one of BLAS, CBLAS, LAPACK, LAPACKE, use the following in the host of the recipe,

    - libblas
    - libcblas
    - liblapack
    - liblapacke


You should specify only the libraries that the package needs. (i.e. if the package doesn’t need LAPACK, remove liblapack and liblapacke)

At recipe build time, above requirements would download the NETLIB’s reference implementations and build your recipe against those. At runtime, by default the following packages will be used.

- openblas   # [not win]
- mkl        # [win]

If a package needs a specific implementation’s internal API for more control you can have,

    - {{ blas_impl }}
    - libblas * *{{ blas_impl }}
    - {{ blas_impl }}

This would give you a matrix builds for different blas implementations. If you only want to support a specific blas implementation,

    - openblas
    - libblas * *openblas
    - openblas


blas_* features should not be used anymore.

Switching BLAS implementation

You can switch your BLAS implementation by doing,

conda install "libblas=*=*mkl"
conda install "libblas=*=*openblas"
conda install "libblas=*=*blis"
conda install "libblas=*=*netlib"

This would change the BLAS implementation without changing the conda packages depending on BLAS.

The following legacy commands are also supported as well.

conda install "blas=*=mkl"
conda install "blas=*=openblas"
conda install "blas=*=blis"
conda install "blas=*=netlib"


If you want to commit to a specific blas implementation, you can prevent conda from switching back by pinning the blas implementation in your environment. To commit to mkl, add blas=*=mkl to <conda-root>/envs/<env-name>/conda-meta/pinned, as described in the conda-docs.

How it works

At recipe build time, the netlib packages are used. This means that the downstream package will link to in the libblas=*=*netlib and will use only the reference implementation’s symbols.

libblas and libcblas versioning is based on the Reference LAPACK versioning which at the time of writing is 3.8.0. Since the BLAS API is stable, a downstream package will only pin to 3.* of libblas and libcblas. On the other hand, liblapack and liblapacke pins to 3.8.*.

In addition to the above netlib package, there are other variants like libblas=*=*openblas, which has openblas as a dependency and has a symlink from to libblas=3.8.0=*openblas pins the openblas dependency to a version that is known to support the BLAS 3.8.0 API. This means that, at install time, the user can select what BLAS implementation they like without any knowledge of the version of the BLAS implementation needed.


matplotlib on conda-forge comes in two parts. The core library is in matplotlib-base. The actual matplotlib package is this core library plus pyqt. Most, if not all, packages that have dependence at runtime on matplotlib should list this dependence as matplotlib-base unless they explicitly need pyqt. The idea is that a user installing matplotlib explicitly would get a full featured installation with pyqt. However, pyqt is a rather large package, so not requiring it indirectly is better for performance. Note that you may need to include a yum_requirements.txt file in your recipe with


if you import parts of matplotlib that link to libX11.

pybind11 ABI Constraints

Sometimes when different python libraries using pybind11 interact via lower-level C++ interfaces, the underlying ABI between the two libraries has to match. To ease this use case, we have a pybind11-abi metapackage that can be used in the host section of a build. Its version is pinned globally and it has a run export on itself, meaning that builds with this package in host will have a runtime constraint on it. Further, the pybind11 has a run constraint on the ABI metapackage to help ensure consistent usage.

To use this package in a build, put it in the host environment like so

    - pybind11-abi

Empty Python packages

For some features introduced in later Python versions, the Python community creates backports, which makes these features available for earlier versions of Python as well. One example here is dataclasses which was introduced with Python3.7 but is available as a backport for Python3.6 too. Therefore, most upstream packages make those backports only mandatory for specific versions of Python and exclude them otherwise.

Implementing this restriction in conda-forge is currently only possible through the use of skips which restricts the corresponding conda-forge recipes from becoming noarch.

Therefore, some conda-forge recipes only create an actual package on specific Python versions and are otherwise an empty placeholder. This allows them to be safely installed under all Python versions and makes using skips unnecessary.

Currently available packages:


Available on:

Empty on:


python >=3.6,<3.7

python >=3.7


python >=3

Noarch builds

Noarch packages are packages that are not architecture specific and therefore only have to be built once.

Declaring these packages as noarch in the build section of the meta.yaml, reduces shared CI resources. Therefore all packages that qualify to be noarch packages, should be declared as such.

Noarch python

The noarch: python directive, in the build section, makes pure-Python packages that only need to be built once.

In order to qualify as a noarch python package, all of the following criteria must be fulfilled:

  • No compiled extensions

  • No post-link or pre-link or pre-unlink scripts

  • No OS-specific build scripts

  • No python version specific requirements

  • No skips except for python version. If the recipe is py3 only, remove skip statement and add version constraint on python in host and run section.

  • 2to3 is not used

  • Scripts argument in is not used

  • If console_scripts entry_points are defined in or setup.cfg, they are also listed in the build section of meta.yaml

  • No activate scripts

  • Not a dependency of conda


While noarch: python does not work with selectors, it does work with version constraints. skip: True  # [py2k] can be replaced with a constrained python version in the host and run subsections: say python >=3 instead of just python.


Only console_scripts entry points have to be listed in meta.yaml. Other entry points do not conflict with noarch and therefore do not require extra treatment.


noarch is a statement about the package’s source code and not its install environment. A package is still considered noarch even if one of its dependencies is not available on a given platform. If this is the case, conda will display a helpful error message describing which dependency couldn’t be found when it tries to install the package. If the dependency is later made available, your package will be installable on that platform without having to make any changes to the feedstock. However, keep in mind that since noarch packages are built on Linux, all dependencies must be available on Linux.

If an existing python package qualifies to be converted to a noarch package, you can request the required changes by opening a new issue and including @conda-forge-admin, please add noarch: python.

Noarch generic


add some information on r packages which make heavy use of noarch: generic

Build matrices

Currently, python, vc, r-base will create a matrix of jobs for each supported version. If python is only a build dependency and not a runtime dependency (eg: build script of the package is written in Python, but the package is not dependent on Python), use build section

Following implies that python is only a build dependency and no Python matrix will be created.

  - python
  - some_other_package

Note that host should be non-empty or compiler jinja syntax used or build/merge_build_host set to True for the build section to be treated as different from host.

Following implies that python is a runtime dependency and a Python matrix for each supported Python version will be created.

  - python

conda-forge.yml’s build matrices is removed in conda-smithy=3. To get a build matrix, create a conda_build_config.yaml file inside the recipe folder. For example, the following will give you 2 builds and you can use the selector vtk_with_osmesa in the meta.yaml

  - False
  - True

You need to rerender the feedstock after this change.

Requiring newer macOS SDKs

conda-forge uses macOS SDK 10.9 to build software so that they can be deployed to all macOS versions newer than 10.9. Sometimes, some packages require a newer SDK to build with. While the default version 10.9 can be overridden using the following changes to the recipe, it should be done as a last resort. Please consult with core team if this is something you think you need.

To use a new SDK, add the following in recipe/conda_build_config.yaml

# Please consult conda-forge/core before doing this
MACOSX_SDK_VERSION:        # [osx and x86_64]
  - "10.12"                # [osx and x86_64]

Note that this should be done if the error you are getting says that a header is not found or a macro is not defined. This will make your package compile with a newer SDK but with 10.9 as the deployment target. WARNING: some packages might use features from 10.12 if you do the above due to buggy symbol availability checks. For example packages looking for clock_gettime will see it as it will be a weak symbol, but the package might not have a codepath to handle the weak symbol, in that case, you need to update the MACOSX_DEPLOYMENT_TARGET as described below.

After increasing the SDK version, if you are getting an error that says that a function is available only for macOS x.x, then do the following in recipe/conda_build_config.yaml,

# Please consult conda-forge/core before doing this
MACOSX_DEPLOYMENT_TARGET:  # [osx and x86_64]
  - "10.12"                # [osx and x86_64]
MACOSX_SDK_VERSION:        # [osx and x86_64]
  - "10.12"                # [osx and x86_64]

In recipe/meta.yaml, add the following to ensure that the user’s system is compatible.

    - __osx >={{ MACOSX_DEPLOYMENT_TARGET|default("10.9") }}  # [osx and x86_64]

Note that this requires conda>=4.8. If you want to support older conda versions the requirement should be changed from run to run_constrained. Note that conda<4.8 will ignore the condition if it’s a run_constrained on __osx.

Newer C++ features with old SDK

The libc++ library uses Clang availability annotations to mark certain symbols as unavailable when targeting versions of macOS that ship with a system libc++ that do not contain them. Clang always assumes that the system libc++ is used.

The conda-forge build infrastructure targets macOS 10.9 and some newer C++ features such as fs::path are marked as unavailable on that platform, so the build aborts:

error: 'path' is unavailable: introduced in macOS 10.15
note: 'path' has been explicitly marked unavailable here
class _LIBCPP_TYPE_VIS path {

However, since conda-forge ships its own (modern) libcxx we can ignore these checks because these symbols are in fact available. To do so, add _LIBCPP_DISABLE_AVAILABILITY to the defines. For example


PyPy builds

To use the PyPy 3.6 or 3.7 builds you can do the following,

conda create -n pypy36  pypy python=3.6
conda create -n pypy37  pypy python=3.7


As of March 8 2020, if you are using defaults as a low priority channel, then you need to use strict channel priority as the metadata in defaults has not been patched yet which allows cpython extension packages to be installed alongside pypy.

conda config --set channel_priority strict

To build your python package for pypy, wait for the bot to send a PR and contact conda-forge/bot team if a PR is not sent after the dependencies have been built.

To add a dependency just for pypy or cpython, do,

    - spam           # [python_impl == 'cpython']
    - ham            # [python_impl == 'pypy']


You’ll need to rerender the feedstocks after making the above change in order for the python_impl variable to be available to conda-build

To skip the pypy builds, do the following,

  skip: True         # [python_impl == 'pypy']

Using setuptools_scm

The Python module setuptools_scm can be used to manage a package’s version automatically from metadata, such as git tags. The package’s version string is thus not specified anywhere in the package, but encoded in it at install-time.

For conda-build this means that setuptools_scm must be included as a host dependency. Additionally, some attention because the metadata is often not available in the sources. There are two options for how to proceed:

  • For Python package also available on PyPI: Use the PyPi tarball as a source, as it will have the metadata encoded (in such a way that setuptools_scm knows how to find it).

  • Specify the environment variable SETUPTOOLS_SCM_PRETEND_VERSION with the version string. If specified this environment variable is the principle source for setuptools_scm. There are two ways how to do this:

    • If you are using build scripts, in specify:


      and in bld.bat specify:


      Whereby you use that PKG_VERSION has been set with the version string, see Environment variables.

    • Otherwise, if you are directly building from meta.yaml, use for example:

        # [...]
          - SETUPTOOLS_SCM_PRETEND_VERSION={{version}}
        script: "{{ PYTHON }} -m pip install . -vv"

Using CentOS 7

To use the newer CentOS 7 sysroot with glibc 2.17 on linux-64, put the following in your build section.

    - {{ compiler('c') }}
    - sysroot_linux-64 2.17  # [linux64]

You also need to use a newer docker image by setting the following in the conda-forge.yml of your recipe and rerendering.

  linux_64: cos7

Finally, note that the aarch64 and ppc64le platforms already use CentOS 7.

CUDA builds

Although the provisioned CI machines do not feature a GPU, Conda-Forge does provide mechanisms to build CUDA-enabled packages. These mechanisms involve several packages:

  • cudatoolkit: The runtime libraries for the CUDA toolkit. This is what end-users will end up installing next to your package.

  • nvcc: Nvidia’s EULA does not allow the redistribution of compilers and drivers. Instead, we provide a wrapper package that locates the CUDA installation in the system. The main role of this package is to set some environment variables (CUDA_HOME, CUDA_PATH, CFLAGS and others), as well as wrapping the real nvcc executable to set some extra command line arguments.

In practice, to enable CUDA on your package, add {{ compiler('cuda') }} to the build section of your requirements and rerender. The matching cudatoolkit will be added to the run requirements automatically.

On Linux, CMake users are required to use ${CMAKE_ARGS} so CMake can find CUDA correctly. For example:

mkdir build && cd build
cmake ${CMAKE_ARGS} ${SRC_DIR}


How is CUDA provided at the system level?

  • On Linux, Nvidia provides official Docker images, which we then adapt to Conda-Forge’s needs.

  • On Windows, the compilers need to be installed for every CI run. This is done through the conda-forge-ci-setup scripts. Do note that the Nvidia executable won’t install the drivers because no GPU is present in the machine.

How is cudatoolkit selected at install time?

Conda exposes the maximum CUDA version supported by the installed Nvidia drivers through a virtual package named __cuda. By default, conda will install the highest version available for the packages involved. To override this behaviour, you can define a CONDA_OVERRIDE_CUDA environment variable. More details in the Conda docs.

Note that prior to v4.8.4, __cuda versions would not be part of the constraints, so you would always get the latest one, regardless the supported CUDA version.

If for some reason you want to install a specific version, you can use:

conda install your-gpu-package cudatoolkit=10.1

Testing the packages

Since the CI machines do not feature a GPU, you won’t be able to test the built packages as part of the conda recipe. That does not mean you can’t test your package locally. To do so:

  1. Enable the Azure artifacts for your feedstock (see here).

  2. Include the test files and requirements in the recipe like this.

  3. Provide the test instructions. Take into account that the GPU tests will fail in the CI run, so you need to ignore them to get the package built and uploaded as an artifact. Example.

  4. Once you have downloaded the artifacts, you will be able to run:

    conda build --test <pkg file>.tar.bz2

Common problems and known issues

nvcuda.dll cannot be found on Windows

The scripts used to install the CUDA Toolkit on Windows cannot provide nvcuda.dll as part of the installation because no GPU is physically present in the CI machines. As a result, you might get linking errors in the postprocessing steps of conda build:

WARNING (arrow-cpp,Library/bin/arrow_cuda.dll): $RPATH/nvcuda.dll not found in packages,
sysroot(s) nor the missing_dso_whitelist.

.. is this binary repackaging?

For now, you will have to add nvcuda.dll to the missing_dso_whitelist:

    - "*/nvcuda.dll"   # [win]

My feedstock is not building old CUDA versions anymore

With the addition of CUDA 11.1 and 11.2, the default build matrix for CUDA versions was trimmed down to versions 10.2, 11.0, 11.1, 11.2.

If you really need it, you can re-add support for 9.2, 10.0 and 10.1. However, this is not recommended. Adding more CUDA versions to the build matrix will dramatically increase the number of jobs and will place a large burden on our CI resources. Only proceed if there’s a known use case for the extra packages.

  1. Download this migration file.

  2. In your feedstock fork, create a new branch and place the migration file under .ci_support/migrations.

  3. Open a PR and re-render. CUDA 9.2, 10.0 and 10.1 will appear in the CI checks now. Merge when ready!

Adding support for a new CUDA version

Providing a new CUDA version involves five repositores:

The steps involved are, roughly:

  1. Add the cudatoolkit packages in cudatoolkit-feedstock.

  2. Submit the version migrator to conda-forge-pinning-feedstock. This will stay open during the following steps.

  3. For Linux, add the corresponding Docker images at docker-images. Copy the migration file manually to .ci_support/migrations. This copy should not specify a timestamp. Comment it out and rerender.

  4. For Windows, add the installer URLs and hashes to the conda-forge-ci-setup script. The migration file must also be manually copied here. Rerender.

  5. Create the new nvcc packages for the new version. Again, manual migration must be added. Rerender.

  6. When everything else has been merged and testing has taken place, consider merging the PR opened at step 2 now so it can apply to all the downstream feedstocks.

Apple Silicon builds

The new Apple M1 processor is the first Apple Silicon supported by conda-forge osx-arm64 builds. For new builds to be available, via cross-compilation, a migration is required for the package and its dependencies. These builds are experimental as many of them are untested.

To request a migration for a particular package and all its dependencies:

  1. Check the feedstock in question to see if there is already an issue or pull request. Opening an issue here is fine, as it might take a couple iterations of the below, especially if many dependencies need to be built as well.

  2. If nothing is under way, look at the current conda-forge-pinning.

  3. If the package is not listed there, make a PR, adding the package name to the end of osx_arm64.txt. The migration bot should start making automated pull requests to the repo and its dependencies.

  4. Within a few hours, the status page should reflect the progress of the package in question, and help you keep track of progress. Help out if you can!

  5. The feedstock maintainers (who very likely do not have an M1) will work to make any changes required to pass continuous intgration. If you have insight into the particular package, please chime in, but most of all be patient and polite.

  6. Once the new builds are available from, please help the maintainers by testing the packages, and reporting back with any problems… but also successes!

Pre-release builds

Recipe maintainers can make pre-release builds available on conda-forge by adding them to the dev or rc label.

The semantics of these labels should generally follow the guidelines that Python itself follows.

  • rc: Beta and Release Candidate (RC). No new features. Bugfix only.

  • dev: Pre-Alpha and Alpha. These are still packages that could see substantial changes between the dev version and the final release.


alpha and beta labels aren’t used. Given the light usage of labels on the conda-forge channel thus far, it seems rather unnecessary to introduce many labels. dev and rc seem like a nice compromise.


Certain packages (for example black) follow a release cycle in which they have never had a non-beta/alpha release. In these cases the conda packages for those do not need to be published to a prerelease label.

Creating a pre-release build

To create a dev or rc package, a PR can be issued into the dev or rc branch of the feedstock. This branch must change the recipe/conda_build_config.yaml file to point to the <package_name>_dev or <package_name>_rc label.

For example, matplotlib rc releases would include:

  - conda-forge matplotlib_rc

If a pre-release build of B depends on a pre-release build of A, then A should have,

  - conda-forge A_rc

while B should have,

  - conda-forge/label/A_rc,conda-forge,defaults
  - conda-forge B_rc

in recipe/conda_build_config.yaml in their respective feedstocks.


A rerender needs to happen for these changes to reflect in CI files.

Installing a pre-release build

Use the following command, but replace PACKAGE_NAME with the package you want to install and replace LABEL with rc or dev:

conda install -c conda-forge/label/PACKAGE_NAME_LABEL -c conda-forge PACKAGE_NAME

For example, let’s install matplotlib from the rc label:

conda install -c conda-forge/label/matplotlib_rc -c conda-forge matplotlib

Pre-release version sorting

If you wish to add numbers to your dev or rc build, you should follow the guidelines put forth by Continuum regarding version sorting in conda. Also see the source code for conda 4.2.13. The tl;dr here is that conda sorts as follows:

< 1.0
< 1.1dev1    # special case 'dev'
< 1.1.0dev1  # special case 'dev'
== 1.1.dev1   # 0 is inserted before string
< 1.1.0rc1
< 1.1.0

So make sure that you tag your package in such a way that the package name that conda-build spits out will sort the package uploaded with an rc label higher than the package uploaded with the dev label.