說明
機器P40*8
操作系統centos 6.x，系統內核版本3.10.0_3-0-0-26，glibc 版本2.12
驅動版本418.39，cuda版本9.2
需要有root權限(安裝nccl2)和下載編譯依賴
另外glibc版本過低，並不滿足nccl2官方要求。
官方推薦升級操作系統版本到centos 7 ，或者使用docker
安裝gcc 4.8.5
wget ftp://ftp.gnu.org/gnu/gcc/gcc-4.8.5/gcc-4.8.5.tar.gz
tar xvf gcc-4.8.5.tar.gz
cd gcc-4.8.5
#下載編譯依賴，如果下載失敗則手動下載
./contrib/download_prerequisites
mkdir build
cd build
../configure --prefix=/home/huxiaoguang/bin/gcc-4.8.5 --disable-multilib
make -j 8
make install
安裝anaconda3
wget https://mirrors.tuna.tsinghua.edu.cn/anaconda/archive/Anaconda3-5.3.1-Linux-x86_64.sh
sh Anaconda3-5.3.1-Linux-x86_64.sh
#自動設置環境變量到.bashrc
#重新登錄使環境變量生效
which python
python -V
conda create -n paddle-1.6
conda env list
conda activate paddle-1.6
#安裝protobuf
pip install protobuf
安裝cmake
wget https://cmake.org/files/v3.5/cmake-3.5.2.tar.gz
tar xvzf cmake-3.5.2.tar.gz
cd cmake-3.5.2
./bootstrap --prefix=/home/huxiaoguang/bin/cmake-3.5.2
make -j 8
make install
安裝git
wget https://mirrors.edge.kernel.org/pub/software/scm/git/git-2.23.0.tar.gz
tar xvf git-2.23.0.tar.gz
cd git-2.23.0
# centos自帶的curl, ssl庫版本過低，需要使用anaconda3的庫版本
export LD_LIBRARY_PATH=/home/huxiaoguang/anaconda3/lib:$LD_LIBRARY_PATH
./configure --prefix=/home/huxiaoguang/bin/git-2.23.0
make
make install
echo "export PATH=/home/huxiaoguang/bin/git-2.23.0/bin:\$PATH" >> ~/.bash_profile
source ~/.bash_profile
which git
git --version
以下步驟不一定需要，通常anaconda的lib版本可滿足git的使用需求。
測試git clone命令，如提示 SSL connect error，則有可能是nss版本過低造成，需要升級nss庫。
因爲Paddle編譯時需要通過git下載依賴的代碼庫，所以需要保持git命令能正常工作。
wget https://ftp.mozilla.org/pub/security/nss/releases/NSS_3_37_1_RTM/src/nss-3.37.1-with-nspr-4.19.tar.gz
tar xvzf nss-3.37.1-with-nspr-4.19.tar.gz
cd nss-3.37.1
export BUILD_OPT=1
export USE_64=1
make nss_build_all
cd ../dist
cp -r Linux3.10_x86_64_cc_glibc_PTH_64_OPT.OBJ /home/huxiaoguang/bin/nss-3.37
安裝cudnn&nccl2
# 解壓後放到特定目錄
wget https://developer.nvidia.com/compute/machine-learning/cudnn/secure/7.6.4.38/Production/9.2_20190923/cudnn-9.2-linux-x64-v7.6.4.38.tgz
wget https://developer.nvidia.com/compute/machine-learning/nccl/secure/v2.4/prod/nccl-repo-rhel7-2.4.8-ga-cuda9.2-1-1.x86_64.rpm
# 查詢安裝目錄
rpm -qpl nccl-repo-rhel7-2.4.8-ga-cuda9.2-1-1.x86_64.rpm
# 無法修改安裝目錄，必須使用root安裝
sudo rpm -ivh nccl-repo-rhel7-2.4.8-ga-cuda9.2-1-1.x86_64.rpm
sudo rpm -qpl /var/nccl-repo-2.4.8-ga-cuda9.2/libnccl-2.4.8-1+cuda9.2.x86_64.rpm
#/usr/lib64/libnccl.so.2
#/usr/lib64/libnccl.so.2.4.8
#/usr/share/doc/libnccl-2.4.8
#/usr/share/doc/libnccl-2.4.8/LICENSE.txt
sudo rpm -ivh /var/nccl-repo-2.4.8-ga-cuda9.2/libnccl-2.4.8-1+cuda9.2.x86_64.rpm
sudo rpm -qpl /var/nccl-repo-2.4.8-ga-cuda9.2/libnccl-devel-2.4.8-1+cuda9.2.x86_64.rpm
#/usr/include/nccl.h
#/usr/lib64/libnccl.so
#/usr/share/doc/libnccl-devel-2.4.8
#/usr/share/doc/libnccl-devel-2.4.8/LICENSE.txt
sudo rpm -ivh /var/nccl-repo-2.4.8-ga-cuda9.2/libnccl-devel-2.4.8-1+cuda9.2.x86_64.rpm
編譯Paddle
git clone https://github.com/PaddlePaddle/Paddle.git
cd paddle
git checkout v1.6.0
mkdir build
cd build
cmake .. -DPY_VERSION=3.7 -DPADDLE_VERSION=1.6.0 -DWITH_GPU=ON -DWITH_TESTING=OFF -DCMAKE_BUILD_TYPE=Release -DCUDNN_ROOT=/home/huxiaoguang/bin/cudnn-v7.6.4/cuda
# 網絡問題可能會造成編譯依賴下載失敗，可以適當調小make時的併發數
make -j 6
# 安裝whl包
cd python/dist
pip install *.whl
# 驗證是否成功
cd ~
python
>>> import paddle
>>> print(paddle.__version__)
# 1.6.0