Scikit-build-core provides a variety of useful variables for your CMakeLists.
You can write CMakeLists that support running inside and outside
scikit-build-core using the ${SKBUILD}
variable. This will be defined to "2"
for scikit-build-core (and "1" for classic scikit-build). You can also detect
the version of scikit-build-core with ${SKBUILD_CORE_VERSION}
.
Scikit-build-core provides several useful variables:
${SKBUILD_PROJECT_NAME}
: The name of the project.${SKBUILD_PROJECT_VERSION}
: The version of the project in a form CMake can use.${SKBUILD_PROJECT_VERSION_FULL}
: The exact version of the project including dev & local suffix.${SKBUILD_STATE}
: The run state, one ofsdist
,wheel
,metadata_wheel
,editable
, ormetadata_editable
.
You can directly use FindPython:
find_package(Python COMPONENTS Interpreter Development.Module REQUIRED)
You always want to find at least Interpreter
and the "Module" component of the
"Development" package. You do not want to find the entire "Development" package,
as that include "Embed" component, which is not always present and is not
related to making Python extension modules.
If you are making a Limited ABI / Stable API package, you'll need the
Development.SABIModule
component instead (CMake 3.26+). You can use the
SKBUILD_LIMITED_API
variable to check to see if it was requested.
:::{warning} :name: soabi
If you want to cross-compile to Windows ARM, you'll need to use
${SKBUILD_SOABI}
, which is always correct, instead of trusting FindPython's
Python_SOABI
value. You can manually set the extension suffix after making a
target:
if(CMAKE_SYSTEM_NAME STREQUAL "Windows")
set_property (TARGET ${name} PROPERTY SUFFIX ".${SKBUILD_SOSABI}.pyd")
else()
set_property (TARGET ${name} PROPERTY SUFFIX ".${SKBUILD_SOSABI}${CMAKE_SHARED_MODULE_SUFFIX}")
endif()
A quicker way to do this would be to instead override Python_SOABI
after
find_package(Python)
:
set(Python_SOABI ${SKBUILD_SOABI})
However, this isn't officially supported upstream, and only works due to the way this variable is used when creating targets.
:::
If you want to use the old, deprecated FindPythonInterp and FindPythonLibs
instead, you can. Though it should be noted that FindPythonLibs requires a trick
to make it work properly if a Python library is not preset (like in manylinux):
you have to set PYTHON_LIBRARY
to something (doesn't matter what) to make it
succeed.
Scikit-build-core includes the site-packages directory in CMake's search path,
so packages can provide a find package config with a name matching the package
name - such as the pybind11
package.
Third party packages can declare entry-points cmake.module
and cmake.prefix
,
and the specified module will be added to CMAKE_PREFIX_PATH
and
CMAKE_MODULE_PATH
, respectively. Currently, the key is not used, but
eventually there might be a way to request or exclude certain entry-points by
key.
Scikit-build-core will install directly into platlib, which will end up in
site-packages. If you are used to scikit-build, you might find targeting
/<module_name>
to be more natural. You can mimic the old behavior with a
configuration option (wheel.install-dir
). However, scikit-build-core is more
powerful, and allows you to install multiple packages or top-level extension
modules if you need to.
You can access all of the possible output directories, regardless of configuration, with the variables:
${SKBUILD_PLATLIB_DIR}
: The original platlib directory. Anything here goes directly to site-packages when a wheel is installed.${SKBUILD_DATA_DIR}
: The data directory. Anything here goes to the root of the environment when a wheel is installed (use with care).${SKBUILD_HEADERS_DIR}
: The header directory. Anything in here gets installed to Python's header directory.${SKBUILD_SCRIPTS_DIR}
: The scripts directory. Anything placed in here will go tobin
(Unix) orScripts
(Windows).${SKBUILD_METADATA_DIR}
: The dist-info directory. Licenses go in thelicenses
subdirectory. _Note that CMake is not run in theprepare_metadata_\*
hooks, so anything written to this directory will only be present when writing wheels._${SKBUILD_NULL_DIR}
: Anything installed here will not be placed in the wheel.
You can activate the limited ABI by setting When you do that,
${SKBUILD_SABI_COMPONENT}
will be set to Development.SABIModule
if you can
target this (new enough CPython), and will remain an empty string otherwise
(PyPy). This allows the following idiom:
find_package(Python REQUIRED COMPONENTS Interpreter Development.Module ${SKBUILD_SABI_COMPONENT})
This will add this only if scikit-build-core is driving the compilation and is
targeting ABI3. If you want to support limited ABI from outside
scikit-build-core, look into the OPTIONAL_COMPONENTS
flag for find_package
.
When defining your module, if you only support the Stable ABI after some point, you should use (for example for 3.11):
if(NOT "${SKBUILD_SABI_COMPONENT}" STREQUAL "")
python_add_library(some_ext MODULE WITH_SOABI USE_SABI 3.11 ...)
else()
python_add_library(some_ext MODULE WITH_SOABI ...)
endif()
This will define Py_LIMITED_API
for you. If you want to support building
directly from CMake, you need to protect this for Python version,
Python_INTERPRETER_ID STREQUAL Python
, and free-threading Python 3.13+ doesn't
support ABI3 either.
If you are using nanobind
's nanobind_add_module
, the STABLE_ABI
flag does
this automatically for you for 3.12+.
Scikit-build-core does not include helpers for F2Py or Cython like scikit-build classic yet. These will be carefully reimagined soon.