xgboost 多gpu支持 编译

xgboost 多gpu支持 编译

Ubuntu 18.04.2
Linux 4.15.0-46-generic
gcc (Ubuntu 7.3.0-27ubuntu1~18.04) 7.3.0

cuda 10.0
    https://docs.nvidia.com/cuda/archive/10.0/cuda-installation-guide-linux/index.html#verify-you-have-supported-version-of-linux
    安装略

nccl2
    git clone https://github.com/NVIDIA/nccl.git
    cd nccl
    make -j src.build

xgboost
    (建议选择稳定版源码编译 如 0.82)

    mkdir xgboost-src

    git clone --recursive https://github.com/dmlc/xgboost.git
    或
    git clone https://github.com/dmlc/xgboost.git
    git submodule init
    git submodule update

    设置版本0.82(!然而最后安装后的版本是0.81)
    git checkout 3f83dcd

    mkdir build
    cd build
    cmake .. -DUSE_CUDA=ON -DUSE_NCCL=ON -DNCCL_ROOT=/xxx/install/nccl-src/nccl/build
    make -j4

    直至出现类似结果
        ...
        Scanning dependencies of target gpuxgboost
        [ 95%] Linking CXX static library libgpuxgboost.a
        [ 95%] Built target gpuxgboost
        Scanning dependencies of target runxgboost
        [ 97%] Building CXX object CMakeFiles/runxgboost.dir/src/cli_main.cc.o
        [ 98%] Linking CXX executable ../xgboost
        [ 98%] Built target runxgboost
        Scanning dependencies of target xgboost
        [100%] Linking CXX shared library ../lib/libxgboost.so
        [100%] Built target xgboost
    cd ../python-package
    python setup.py install

备注: 如果切换 使用 update-alternatives gcc/g++ 版本时，可能会出现各种引用异常，此时建议切换到gcc/g++某个已安装版本(如7.3)， 重启机器

------------------------------------------------------
tensorflow (对应cuda 10.0)
     tensorflow-gpu        1.13.1
     pip install tensorflow-gpu

------------------------------------------------------
torch  (对应cuda 10.0)
    pip install https://download.pytorch.org/whl/cu100/torch-1.0.1.post2-cp36-cp36m-linux_x86_64.whl
    pip install torchvision