从源代码构建#

以下说明适用于希望从源代码构建 wholegraph 的用户。这些说明已在支持的 Linux、CUDA 和 Python 发行版上进行测试 - 有关支持环境的列表，请参阅 RAPIDS 入门指南。其他操作系统可能兼容，但目前尚未测试。

wholegraph 包包含 C/C++ CUDA 部分和 Python 部分。需要安装这两个库，cuGraph 才能正常运行。C/C++ CUDA 库是 libwholegraph，Python 库是 pylibwholegraph。

先决条件#

编译器:

  • gcc 版本 11.0+

  • nvcc 版本 11.0+

  • cmake 版本 3.26.4+

CUDA:

  • CUDA 11.8+

  • Volta 架构或更高版本

您可以从 https://developer.nvidia.com/cuda-downloads 获取 CUDA。

其他软件包:

  • ninja

  • nccl

  • cython

  • setuputils3

  • scikit-learn

  • scikit-build-core

  • nanobind>=0.2.0

构建 wholegraph#

要从源代码安装 wholegraph，请确保满足依赖项。

克隆仓库并配置 Conda 环境#

GIT 克隆仓库的某个版本

# Set the location to wholegraph in an environment variable WHOLEGRAPH_HOME
export WHOLEGRAPH_HOME=$(pwd)/wholegraph

# Download the wholegraph repo - if you have a forked version, use that path here instead
git clone https://github.com/rapidsai/wholegraph.git $WHOLEGRAPH_HOME

cd $WHOLEGRAPH_HOME

创建 conda 开发环境

# create the conda environment (assuming in base `wholegraph` directory)

# for CUDA 11.x
conda env create --name wholegraph_dev --file conda/environments/all_cuda-118_arch-x86_64.yaml

# activate the environment
conda activate wholegraph_dev

# to deactivate an environment
conda deactivate

  • 随着开发包含/更改依赖项，环境可以更新。为此，运行

# Where XXX is the CUDA version
conda env update --name wholegraph_dev --file conda/environments/all_cuda-XXX_arch-x86_64.yaml

conda activate wholegraph_dev

使用 build.sh 脚本构建和安装#

使用 build.sh 脚本可以轻松编译和安装 wholegraph。要构建和安装，只需执行

$ cd $WHOLEGRAPH_HOME
$ ./build.sh clean
$ ./build.sh libwholegraph
$ ./build.sh pylibwholegraph

构建脚本上还有其他几个选项可供高级用户使用。build.sh 选项

build.sh [<target> ...] [<flag> ...]
 where <target> is:
   clean                    - remove all existing build artifacts and configuration (start over).
   uninstall                - uninstall libwholegraph and pylibwholegraph from a prior build/install (see also -n)
   libwholegraph            - build the libwholegraph C++ library.
   pylibwholegraph          - build the pylibwholegraph Python package.
   tests                    - build the C++ (OPG) tests.
   benchmarks               - build benchmarks.
   docs                     - build the docs
 and <flag> is:
   -v                          - verbose build mode
   -g                          - build for debug
   -n                          - no install step
   --allgpuarch               - build for all supported GPU architectures
   --cmake-args=\\\"<args>\\\" - add arbitrary CMake arguments to any cmake call
   --compile-cmd               - only output compile commands (invoke CMake without build)
   --clean                    - clean an individual target (note: to do a complete rebuild, use the clean target described above)
   -h | --h[elp]               - print this text

 default action (no args) is to build and install 'libwholegraph' then 'pylibwholegraph' targets

examples:
$ ./build.sh clean                        # remove prior build artifacts (start over)
$ ./build.sh

# make parallelism options can also be defined: Example build jobs using 4 threads (make -j4)
$ PARALLEL_LEVEL=4 ./build.sh libwholegraph

Note that the libraries will be installed to the location set in `$PREFIX` if set (i.e. `export PREFIX=/install/path`), otherwise to `$CONDA_PREFIX`.

独立构建每个部分#

构建和安装 C++/CUDA libwholegraph 库#

CMake 依赖于 nvcc 可执行文件位于您的路径中或定义在 $CUDACXX 中。

本项目使用 cmake 构建 C/C++ 库。要配置 cmake，运行

# Set the location to wholegraph in an environment variable WHOLEGRAPH_HOME
export WHOLEGRAPH_HOME=$(pwd)/wholegraph

cd $WHOLEGRAPH_HOME
cd cpp                                        # enter cpp directory
mkdir build                                   # create build directory
cd build                                      # enter the build directory
cmake .. -DCMAKE_INSTALL_PREFIX=$CONDA_PREFIX

# now build the code
make -j                                       # "-j" starts multiple threads
make install                                  # install the libraries

默认安装位置分别是 $CMAKE_INSTALL_PREFIX/lib 和 $CMAKE_INSTALL_PREFIX/include/wholegraph。

构建和安装 Python 包#

将 Python 包构建并安装到您的 Python 路径

cd $WHOLEGRAPH_HOME
cd python
cd pylibwholegraph
python setup.py build_ext --inplace
python setup.py install    # install pylibwholegraph

运行测试#

使用数据集运行 C++ 或 Python 测试

  • 带数据集的 Python 测试

    cd $WHOLEGRAPH_HOME
cd python
pytest

  • C++ 独立测试

    从构建目录

    # Run the tests
cd $WHOLEGRAPH_HOME
cd cpp/build
gtests/PARALLEL_UTILS_TESTS		# this is an executable file

注意：此 conda 安装仅适用于 Linux 以及 Python 版本 3.10、3.11 和 3.12。

创建文档#

Python API 文档可以从 ./docs/wholegraph 目录生成。或者通过使用“./build.sh docs”生成。

归属#

部分内容改编自 https://github.com/pytorch/pytorch/blob/master/CONTRIBUTING.md