系统：redhat el7 同样适用于centos7

nvidia显卡驱动  版本:     410.129    CUDA Toolkit:     10.0

cuda 10.0

cudnn 7.6.5

python 3.7

tensorflow-gpu 1.14.0

第一步安装anaconda

bash Anaconda3-2020.02-Linux-x86_64.sh

vim /etc/profile

添加

#Anaconda

export PATH=/root/anaconda3/bin:$PATH

source /etc/profile

查看python版本

python --version

第二步安装nvidia驱动

1、禁止Linux系统自带的集成驱动

vim /etc/modprobe.d/blacklist.conf

使用vim在打开的文件后面添加几行：

blacklist vga16fb

blacklist nouveau

blacklist rivafb

blacklist rivatv

blacklist nvidiafb

chmod 666 /etc/modprobe.d/blacklist.conf

之后，重启服务器 reboot

重启之后执行：lsmod | grep nouveau

2.安装依赖包

在安装驱动的时候需要几个依赖包，比如gcc，g++，make

yum -y install gcc kernel-devel kernel-headers g++ make

去https://www.nvidia.cn/Download/index.aspx?lang=cn找到自己对应系统的显卡驱动

Tesla Driver for Linux RHEL 7

版本:     410.129

发布日期:     2019.9.4

操作系统:     Linux 64-bit RHEL7

CUDA Toolkit:     10.0

语言:     Chinese (Simplified)

文件大小:     111.5 MB

3、驱动安装  CUDA Toolkit:     10.0

rpm -ivh nvidia-diag-driver-local-repo-rhel7-410.129-1.0-1.x86_64.rpm

yum install -y cuda-drivers

执行如下命令，查询driver版本。

rpm -qa | grep -i nvidia

第三步，安装cuda10.0

1、如果下载的是sh：

sudo sh cuda_10.0.130_410.48_linux.run

有一个地方需要注意：在询问是否安装gpu drivers时要选择no

Install NVIDIA Accelerated Graphics Driver for Linux-x86_64 410.48?

(y)es/(n)o/(q)uit: n

！！！！卸载sh 重装的时候用

！！！！cd /usr/local/cuda/bin

！！！！sudo ./uninstall_cuda_10.0.pl

reboot

查看

nvidia-smi

2、配置环境变量

vim ~/.bashrc

export PATH="/usr/local/cuda/bin:$PATH"

export LD_LIBRARY_PATH="/usr/local/cuda/lib64:$LD_LIBRARY_PATH"

source ~/.bashrc

vim /etc/profile

export PATH=/usr/local/cuda/bin:$PATH

export CUDA_INSTALL_PATH=/usr/local/cuda

export LD_LIBRARY_PATH=$CUDA_INSTALL_PATH/lib64:$LD_LIBRARY_PATH

source /etc/profile

3、创建链接文件：

vim /etc/ld.so.conf.d/cuda.conf

添加如下语句：

/usr/local/cuda/lib64

然后执行

sudo ldconfig

第四步，解压cudnn-10.0-7.6.5

tar -zxvf cudnn-10.0-linux-x64-v7.6.5.326.tgz

cp /data/software/cuda/include/cudnn.h /usr/local/cuda/include

cp /data/software/cuda/lib64/libcudnn* /usr/local/cuda/lib64

chmod a+r /usr/local/cuda/include/cudnn.h /usr/local/cuda/lib64/libcudnn*

更新一下链接

cd /usr/local/cuda/lib64

sudo ln -sf libcudnn.so.7.6.5 libcudnn.so.7

sudo ln -sf libcudnn.so.7 libcudnn.so

sudo ldconfig

查看一下nvcc的信息验证安装是否成功。

nvcc -V

nvcc --version

卸载

rpm -qa | grep -i nvidia

yum remove nvidia-driver-local-repo-rhel7-440.64.00-1.0-1.x86_64

rpm -qa | grep -i cuda

yum remove cuda-repo-rhel7-10-2-local-10.2.89-440.33.01-1.0-1.x86_64

yum clean all

运行实例代码报错

ImportError: /lib64/libm.so.6: version `GLIBC_2.23' not found (required by /root/anaconda3/lib/python3.7/site-packages/tensorflow/python/_pywrap_tensorflow_internal.so)

1. glibc下载

从http://www.gnu.org/software/libc/ 下载源代码

2. 安装

tar -zxvf glibc-2.23.tar.gz

cd glibc-2.23

mkdir build

cd build

echo $LD_LIBRARY_PATH

../configure --prefix=/data/glibc-2.23    (../configure --prefix=/usr --disable-profile --enable-kernel=2.6.32 --enable-obsolete-rpc)

报错 LD_LIBRARY_PATH shouldn't contain the current directory

方法：临时修改 vim ~/.bashrc

LD_LIBRARY_PATH=

(/usr/local/cuda-10.0/lib64:

export LD_LIBRARY_PATH=

echo $LD_LIBRARY_PATH)

编译后记得修改回来

make -j4

make install

继续报错

gawk: error while loading shared libraries: /lib64/libm.so.6: invalid ELF header

make[2]: *** [/data/software/glibc-2.23/build/math/stubs] Error 127

make[2]: Leaving directory `/data/software/glibc-2.23/math'

make[1]: *** [math/subdir_install] Error 2

make[1]: Leaving directory `/data/software/glibc-2.23'

make: *** [install] Error 2

解决方法

cd /lib64

ls -l | grep libm

发现有新的libm-2.23.so文件生成，这个时候更改了一下软链：

unlink libm.so.6

ln -s libm-2.23.so libm.so.6

继续make install