Native C++ SDK archives

Download sdk-* artifacts from a successful CI run; releases attach the archives and checksums. Manual publishing-workflow runs produce CI artifacts without uploading to PyPI or a release.

Extract an SDK archive and point CMake at its directory:

find_package(apxchol CONFIG REQUIRED)
target_link_libraries(my_program PRIVATE apxchol::core)

cmake -S my-program -B build -DCMAKE_PREFIX_PATH=/path/to/extracted-sdk
cmake --build build

The archive contains the shared native library, public headers, Eigen headers, CMake package files and bundled-code licenses. Use the imported CMake target to retain the required C++23 setting, index layout and backend definitions. SDKs use 32-bit indices and portable tuning: x86-64-v2 on Linux, arm64 on macOS.

Compiler and runtime requirements

Use a compatible C++23 compiler and standard library. The compiler’s OpenMP runtime family must match the SDK; CMake rejects a different family. Linux archives target glibc 2.28+ and use GCC 14 with GNU OpenMP (libgomp). macOS builds use Clang, libc++ and arm64 LLVM OpenMP (libomp). Install the matching OpenMP development package when it is not supplied by the compiler. The zlib runtime is also required.

A CPU SDK does not need a GPU toolkit. A CUDA SDK includes CPU and CUDA routes and currently requires CUDA Toolkit 13.3 or newer when compiling consumers, including CPU-route consumers: its public headers and exported CMake target use CUDA. Use a compatible NVIDIA driver to execute GPU routes. A Metal SDK includes CPU and Metal routes and needs Apple’s system frameworks, plus libomp. The macOS deployment target is 15.0. The release set has Linux CPU, Linux CPU+CUDA and macOS CPU+Metal archives; each uses one shared-library and index-layout configuration. Use the headers and library from the same archive.

Keep the shared library accessible at runtime when installing your program. CMake handles the build-tree library path; set your application’s installation RPATH or install the SDK library into its normal library search path.

Building an archive

From the repository root, use the existing native install rules:

bash .github/scripts/build-sdk.sh CPU

Use CUDA on Linux or METAL on macOS for the GPU-enabled variants. CUDA uses all-major, matching the wheel build; additional arguments are ordinary CMake arguments. The CUDA archive producer uses the same CUDA 13.4 RPM packages and license paths as the Linux wheel build. Each archive has a SHA256 checksum. The script creates a new staging prefix and refuses to overwrite an existing one.