https://tensorflow.google.cn/install/source#gpu_support_2
更新docker内cuda、cudnn正确方式:
一、拉取官方镜像
docker拉取官方镜像(默认没有显卡驱动):
- nvidia-docker pull nvidia/cuda:9.0-cudnn7-devel-ubuntu16.04
可选择镜像:https://hub.docker.com/layers/nvidia/cuda/9.0-cudnn7-devel-ubuntu16.04/images/sha256-5b4a2ecdf7e7594e8f6fc41c0ed0e6a716bfb6ddc0e4852a3dde58eae4692d73
二、安装docker镜像显卡驱动
apt-get update
apt-get install software-properties-common python-software-properties
add-apt-repository ppa:graphics-drivers/ppa
apt-get install ubuntu-drivers-common
apt update
ubuntu-drivers devices
apt install nvidia-384
或 ubuntu-drivers autoinstall
或手动.run安装 chomx a+x 驱动 && ./驱动
ctrl+p+q退出
nvidia-docker commit hash_value img_name #保存镜像
推荐./run安装,以保证host和docker镜像内部驱动版本一致
驱动下载网站:https://www.nvidia.com/Download/driverResults.aspx/143548/en-us
.run安装:https://blog.csdn.net/u014682691/article/details/80605201
参考资料:https://onlycaptain.github.io/2018/08/18/Ubuntu%E4%B8%8BNvidia%E9%A9%B1%E5%8A%A8%E5%AE%89%E8%A3%85/
三、安装Host主机显卡驱动
同步骤二
四、重启主机
五、启动docker
dockerd& #docker守护进程
nvidia-docker-plugin& #nvidia-docker守护进程
service docker start
service nvidia-docker start
六、测试
测试Host主机:
nvidia-smi
更新驱动版本后需更新volume:
nvidia-docker volume create nvidia_driver_390.46
测试镜像:
docker运行:
nvidia-docker run -it img_name /bin/bash
进入docker后:
nvidia-smi
nvcc --version
其他可能遇到的问题
挂载Nvidia驱动:
modprobe nvidia
查看驱动版本:
cat /proc/driver/nvidia/version
dpkg error问题处理:
https://www.cnblogs.com/ilym/p/8387561.html
https://blog.csdn.net/qingfengxiaosong/article/details/87889995
Ubuntu中Failed to initialize NVML: Driver/library version mismatch问题的解决:
https://blog.csdn.net/qq_40200387/article/details/90341107