tensorflow-gpu & Nvidia GPU & Cuda & Cudnn 环境配置

Contents

• Install Nvidia GPU
• Install CUDA
• Install cuDNN
• Install tensorflow-gpu
• Issues

---

Ubuntu 16.04 配置版本

• Ubuntu Kernel Version 4.10.09

• GeForce GTX 1080 Ti

• NVIDIA 390.87

• cuda 9.0

• cuDNN v7.1.4

• tensorflow-gpu 1.10.0

• python 3.5

---

• Ubuntu Kernel Version 4.15.0

• GeForce RTX 2080 Ti

• NVIDIA 410.48

• cuda 10.0

• cuDNN v7.5.0

• tensorflow-gpu 1.13.1

• python 3.5

---

0. 更新软件源 & 确认配置版本 - Update and Upgrade

$ sudo apt-get update
$ sudo apt-get upgrade

(Kernel Version 4.10.09)

Install Version: NVIDIA-Linux-x86_64-390.87.run

Download Nvidia Drivers: [Nvidia Link]

---

1. 卸载所有原驱动 - uninstall all original drivers for safety

$ sudo apt-get purge nvidia*

---

2. 禁用nouveau - Disabling Nouvea

• 新建-blacklist-nouveau.conf 输⼊指令:(Create a blacklist file)

$ sudo vim /etc/modprobe.d/blacklist-nouveau.conf

• 往文件中写⼊-input :(with the following contents)

blacklist nouveau
options nouveau modeset=0

禁 Ubuntu自带开源驱动nouveau，写入后 After above 重启系统 - Disabling Nouveau

$ sudo reboot 

在终端执行行命令:(Terminal)

$ lsmod | grep nouveau

查看nouveau模块是否被加载，若无输出，则执行下一步 - If no output go to the next step.

---

3. 安装Nvidia驱动 - Install the Drivers

Ctrl + Alt + F1-( Enter virtual consoles )进入tty1命令行界面 
Ctrl + Alt + F7-( Return back to GUI )回到桌面系统界面

禁⽤X服务 - Kill your current X server session by typing

$ sudo service lightdm stop

给驱动run文件赋予执行权限(根据版本执行 - Check your NVIDIA version)

$ sudo chmod a+x NVIDIA-Linux-x86_64-390.87.run
$ sudo ./NVIDIA-Linux-x86_64-390.87.run -no-opengl-files -no-x-check -no-nouveau-check

[Tips] (登录界面循环问题 - Login loop issue)

• -no-opengl-files 只安装驱动文件，不安装OpenGL文件 (no install OpenGL file)
• -no-x-check 安装驱动时不检查X服务器 (no check X server)
• -no-nouveau-check 安装驱动时不检查nouveau模块 （no check nouveau module）

安装驱动时 (Downloading)

“Would you like to run the nvidia-xconfig utility to automatically update your X configuration file...”
Choose No.
After above: $sudo reboot

---

4. Nvidia驱动安装完成 - Check Driver was successfully installed

$ nvidia-smi

---

5. 安装CUDA - CUDA Toolkit 9.0 Downloads

Download Installer for Linux Ubuntu 16.04 x86_64

Download CUDA: cuda_9.0.176_384.81_linux.run [CUDA Link]

Install CUDA

$ sudo ./cuda_9.0.176_384.81_linux.run --no-opengl-libs
...

[accept] #同意安装
[n]      #安装Driver，将自动安装CUDA版本相匹配的Nvidia驱动, - No install Driver
[y]      #安装CUDA Toolkit install
<Enter>  #安装到默认目录
[y]      #创建安装目录的软链接
[n]      #不复制Samples，因为在安装目录下有/samples

A. vim 打开.bashrc 在末行加⼊以下命令 - Add following lines to .bashrc

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

执行指令更新 .bashrc 文件 - Reload .bashrc with

$ source .bashrc

B. 安装及路径测试，查看CUDA版本 CUDA Sample Testing:

$ nvcc -V

编译并测试设备 deviceQuery:

$ cd /usr/local/cuda-9.0/samples/1_Utilities/deviceQuery
$ sudo make
$ ./deviceQuery

编译并测试带宽 bandwidthTest:

$ cd ../bandwidthTest
$ sudo make
$ ./bandwidthTest

如果两个测试的结果都是 Result = PASS CUDA 安装成功 - Install successed

 $ cd /usr/local/cuda-9.0/samples/1_Utilities/deviceQuery
 $ sudo make
 $ ./deviceQuery
 $ cd ../bandwidthTest
 $ sudo make
 $ ./bandwidthTest

---

6. 安装cuDNN - Download cuDNN

Download Version: cuDNN v7.1.4 (May 16, 2018), for CUDA 9.0 [cuDNN Link]

解压后的 cudnn-9.0-linux-x64-v7.1.tgz ⽂文件cuda，执行以下指令安装:-Install cudnn

$ tar -zxvf cudnn-9.0-linux-x64-v7.1.tgz
$ cd cuda 
$ sudo cp lib64/lib* /usr/local/cuda/lib64/
$ sudo cp include/cudnn.h /usr/local/cuda/include/

然后更新网络连接:-Updata Network Connection

$ cd /usr/local/cuda/lib64/ 
$ sudo chmod +r libcudnn.so.7.1.4
$ sudo ln -sf libcudnn.so.7.1.4 libcudnn.so.7
$ sudo ln -sf libcudnn.so.7 libcudnn.so 
$ sudo ldconfig

---

7. 安装Tensorflow-gpu - Install Tensorflow-gpu

注意这里 tensorflow-gpu 的版本 cuda 会报错误

CUDA 9.0 - tensorflow-gpu==1.10.0
CUDA 10.0 - tensorflow-gpu==1.13.1

$ sudo pip3 uninstall tensorflow
$ pip3 install --user tensorflow-gpu==1.10.0

---

8. 测试 - Testing

---

9. 常见配置问题 - Issues

登⼊界⾯死循环问题: Login loop issue

1. 进入文本界面: CTRL+ALT+F1

2. Uninstall any previous drivers: - 删除 nvidia 相关文件

$ sudo apt-get remove nvidia-*
$ sudo apt-get autoremove       

3. Uninstall the drivers from the .run file: - 卸载 nvidia 驱动

$ sudo nvidia-uninstall

4. 此时，重启 - Reboot -> login normally

5. 驱动重新安装 - Driver reinstall

---

NVIDIA-SMI has failed because it couldn't communicate with the NVIDIA driver. Make sure that the latest NVIDIA driver is installed and running issue

0. 确认 Kernel 版本是否高于 4.10

$ uname -a

#目前使用版本为 4.15
Linux CAI 4.15.0-50-generic #54~16.04.1-Ubuntu SMP Wed May 8 15:55:19 UTC 2019 x86_64 x86_64 x86_64 GNU/Linux

若版本高于 4.10 必须升级， 降级方法如下

wget http://kernel.ubuntu.com/~kernel-ppa/mainline/v4.15.7/linux-headers-4.15.7-041507_4.15.7-041507.201802280530_all.deb

wget http://kernel.ubuntu.com/~kernel-ppa/mainline/v4.15.7/linux-headers-4.15.7-041507-generic_4.15.7-041507.201802280530_amd64.deb

wget http://kernel.ubuntu.com/~kernel-ppa/mainline/v4.15.7/linux-image-4.15.7-041507-generic_4.15.7-041507.201802280530_amd64.deb

sudo dpkg -i *.deb

升级完成后 nvidia-smi 出现 GPU 使用狀況栏可不用重新安装 Driver， 若未出现可按步骤重新安装 Driver

1. 确认是否插入显卡

$ lspci | grep 'VGA'

#找到卡后，显示显卡讯息
01:00.0 VGA compatible controller: NVIDIA Corporation Device 1b06 (rev a1)

2. 确认 security boot 是否为disable的状态

a. 开机后, 进入Bios 设定画面(若是Acer的电脑, 按Del 或是F2 即可进入Bios)
b. 改成disable 后, 重新开机

---

调整屏幕分辨率: Display resolution issue

METHOD1:

1.添加 /etc/X11/xorg.conf 文件，将此模式保存为默认分辨率。

$ sudo gedit /etc/X11/xorg.conf

2.粘贴以下内容: - copy below -> paste to xorg.conf

Section "Monitor"
Identifier "Configured Monitor"
Modeline "1920x1080_60.00" 173.00 1920 2048 2248 2576 1080
1083 1088 1120 -hsync +vsync
Option "PreferredMode" "1920x1080_60.00"
EndSection
Section "Screen"
Identifier "Default Screen"
Monitor "Configured Monitor"
Device "Configured Video Device"
EndSection
Section "Device"
Identifier "Configured Video Device"
EndSection

METHOD2:

1.生成指定分辨率 （one-off）

$ cvt 1920 1080
# 1920x1080 59.96 Hz (CVT 2.07M9) hsync: 67.16 kHz; pclk: 173.00 MHz
Modeline "1920x1080_60.00"  173.00  1920 2048 2248 2576  1080 1083 1088 1120 -hsync +vsync

2.使用xrandr创建new mode （make newmode）

$ sudo xrandr --newmode "1920x1080_60.00"  173.00  1920 2048 2248 2576  1080 1083 1088 1120 -hsync +vsync

3.添加newmode,终端输入xrand查看显示器名称 (add mode)

$ sudo xrand --addmode [THE NAME OF YOUR DISPLAY] "1920x1080_60.00"

4.将分辨率应用到输出设备 （output display）

$ sudo xrand --output [THE NAME OF YOUR DISPLAY] --mode "1920x1080_60.00"

---

Uninstall - 卸载

卸载 nvidia - uninstall nvidia

$ sudo apt-get purge nvidia*

卸载 Cuda - uninstall Cuda:

$ cd /usr/local/cuda/bin 
$ sudo ./uninstall_cuda_7.5.pl

卸载 Cudnn - uninstall Cudnn:

$ sudo rm -rf /usr/local/cuda/include/cudnn.h
$ sudo rm -rf /usr/local/cuda/lib64/libcudnn

国内 pip 速度问题换源 （阿里云pip源）

cd ~
mkdir .pip
sudo vim .pip/pip.conf

写入：

[global]
index-url = http://mirrors.aliyun.com/pypi/simple/
[install]
trusted-host = mirrors.aliyun.com