环境 ubuntu
一般根据显卡,先安装nvidia-driver,再根据驱动版本,选择cuda、cudnn版本,以及tensorflow、python版本
可以参考:
- https://tensorflow.google.cn/install/source#linux
- https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html
一、安装nvidia-driver
1、下载安装包安装
查找最新显卡驱动
https://www.nvidia.cn/Download/index.aspx?lang=cn
安装
sudo apt-get remove nvidia-*
sudo chmod +x NVIDIA-Linux-x86_64-xxx.xx.run
sudo ./NVIDIA-Linux-x86_64-xxx.xx.run
sudo reboot
nvidia-smi
2、直接apt-get安装
添加源
sudo add-apt-repository ppa:graphics-drivers/ppa
sudo apt update
查看可以安装版本,并安装推荐版本
ubuntu-drivers devices
sudo apt install nvidia-driver-XXX
3、DEBUG:
- 更新nvidia-driver版本时,无法nvidia-smi
$ nvidia-smi
Failed to initialize NVML: Driver/library version mismatch
尝试:
https://blog.youkuaiyun.com/zywvvd/article/details/115500412
之后尝试再次安装nvidia-driver时可能会遇到‘nvidia-drm‘ appears to already be loaded in your kernel…
尝试删干净再装:
https://blog.youkuaiyun.com/u010087338/article/details/107585801
- 安装过程需要禁用nouveau
尝试
https://zhuanlan.zhihu.com/p/439211710
二、安装cuda
-
建议选择runfile,之后添加权限后再进行安装
chmod +x cuda_*_linux.run
sudo ./cuda_*_linux.run
注意若已经安装nvidia-driver驱动后,记得在安装过程中不要重复安装,如下图,第一个x要去掉
- 添加路径
之后在用户目录下.bashrc中最后添加语句指定cuda版本,并source激活
若要对所有用户适用,则在/etc/profile中添加
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda-xx.x/lib64
export PATH=$PATH:/usr/local/cuda-xx.x/bin
export CUDA_HOME=$CUDA_HOME:/usr/local/cuda-xx.x
- nvcc -V验证
三、安装cudnn
- 下载指定版本
https://developer.nvidia.com/rdp/cudnn-download
- 解压、安装,如图
tar -xvf cudnn-linux-x86_64-8.x.x.x_cudaX.Y-archive.tar.xz
sudo cp cudnn-*-archive/include/cudnn*.h /usr/local/cuda/include
sudo cp -P cudnn-*-archive/lib/libcudnn* /usr/local/cuda/lib64
- 赋予权限
sudo chmod a+r /usr/local/cuda-xx.x/include/cudnn*.h /usr/local/cuda-xx.x/lib64/libcudnn*
- 查看安装是否成功
cat /usr/local/cuda-xx.x/include/cudnn.h | grep CUDNN_MAJOR -A 2
cat /usr/local/cuda-xx.x/include/cudnn_version.h | grep CUDNN_MAJOR -A 2
出现上图时成功
参考:
https://blog.youkuaiyun.com/qq_35544714/article/details/109296038