ggml.cpp-build

This repository builds ggml as a shared-library SDK that can be published as CI artifacts and consumed by third-party CMake projects.

Build An SDK

Use a clean build directory, then install into a staging prefix. The staged tree is what should be archived and uploaded to your release page.

cmake -S . -B build ^
  -DBUILD_SHARED_LIBS=ON ^
  -DGGML_BACKEND_DL=ON ^
  -DGGML_CPU=ON ^
  -DGGML_CPU_ALL_VARIANTS=ON ^
  -DGGML_BUILD_TESTS=OFF ^
  -DGGML_BUILD_EXAMPLES=OFF

cmake --build build --config Release
cmake --install build- --config Release --prefix

Package the contents of ggml directly:

• ggml-linux-x86_64.tar.gz
• ggml-macos-arm64.tar.gz
• ggml-windows-x86_64.zip

Each archive should contain the installed SDK layout:

ggml-<platform>/
  manifests.json
  include/
  lib/
  bin/
  lib/cmake/ggml/ggml-config.cmake

The Linux and Windows assemble jobs also publish a sibling *.manifests.json file next to the archive in CI and Releases. It describes the platform, archive format, included backend overlays, the component -> files ownership mapping, and the exact file list so third-party tooling can identify the package before merging it into a larger SDK.

Consume The SDK

The helper module cmake/fetch_prebuilt_ggml.cmake lets a third-party project either:

• use an already unpacked SDK directory
• or download the correct archive for the current platform from your release page

Minimal consumer example:

list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_LIST_DIR}/cmake")
include(fetch_prebuilt_ggml)

set(GGML_ROOT_DIR "" CACHE PATH "Path to an unpacked ggml SDK")
set(GGML_RELEASE_BASE_URL "https://github.com/seasonjs/ggml.cpp-build/releases/download" CACHE STRING "")
set(GGML_RELEASE_TAG "v0.9.8" CACHE STRING "")

if (GGML_ROOT_DIR)
    ggml_import_prebuilt(
        ROOT_DIR "${GGML_ROOT_DIR}"
        OUT_ROOT_DIR GGML_PACKAGE_ROOT
    )
else()
    ggml_import_prebuilt(
        RELEASE_BASE_URL "${GGML_RELEASE_BASE_URL}"
        RELEASE_TAG "${GGML_RELEASE_TAG}"
        OUT_ROOT_DIR GGML_PACKAGE_ROOT
    )
endif()

add_executable(main_app main.cpp)
target_link_libraries(main_app PRIVATE ggml::ggml)

ggml_copy_runtime_binaries(main_app
    ROOT_DIR "${GGML_PACKAGE_ROOT}"
)

The imported targets now propagate the installed include/ directory, so linking ggml::ggml is enough to make #include <ggml.h> work.

ggml_copy_runtime_binaries(...) matters because with BUILD_SHARED_LIBS=ON and GGML_BACKEND_DL=ON, the executable needs the runtime files from bin/ beside it so ggml_backend_load_all() can load the backend plugins.

The local example project at example/CMakeLists.txt demonstrates the same flow.