0. 卸载环境

1、卸载显卡驱动的命令：
假如安装的是NVIDIA-Linux-x86_64-390.116.run：
则运行如下命令：sh NVIDIA-Linux-x86_64-390.116.run --uninstall

2、卸载cuda和cudnn

cd /usr/local/cuda-9.2/bin/
./uninstall_cuda_9.2.pl
rm -rf /usr/local/cuda-9.2/

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

卸载完后需要reboot。

1. 安装GPU驱动

在这里插入图片描述
如果是更早版本，可以在阿里云进行下载，命令为

wget http://mirrors.cloud.aliyuncs.com/opsx/ecs/linux/binary/nvidia/driver/NVIDIA-Linux-x86_64-390.116.run
sh NVIDIA-Linux-x86_64-390.116.run

如果是比较新的版本，需要去官网查询下载https://cn.download.nvidia.cn/tesla：

wget http://cn.download.nvidia.com/tesla/396.82/NVIDIA-Linux-x86_64-396.82.run
sh NVIDIA-Linux-x86_64-396.82.run

使用nvidia-smi查看驱动是否安装成功。

2. 安装cuda和cudnn

由于最终安装的是pytorch-1.4.0-py3.6_cuda9.2.148_cudnn7.6.3_0.tar.bz2(https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch/linux-64/py3.6_cuda9.2.148_cudnn7.6.3_0.tar.bz2)

wget http://mirrors.cloud.aliyuncs.com/opsx/ecs/linux/binary/nvidia/cuda/9.2.148/
cuda_9.2.148.1_linux cuda_9.2.148_396.37_linux                        
wget http://mirrors.cloud.aliyuncs.com/opsx/ecs/linux/binary/nvidia/cuda/9.2.148/cuda_9.2.148.1_linux
wget https://developer.download.nvidia.com/compute/machine-learning/cudn /secure/7.6.3.30/Production/9.2_20190822/cudnn-9.2-linux-x64-v7.6.3.30.tgz?vcg3oGnO2NK5a53HqN1zGD14tDO3RkB0jFuHdq6lAsg_7rX-sM-CAi6OW2_bV7tbCpNprKdqRmW8gtsd3ERtCytItX4p6G-WjxkSYOpEkeiVxxpBe9HCijmQSNkOrcQ7NhbXbyo3udR1CKbihWzNKFgn4k7WMvYq86cJVGVWiErtg-1kX_q5UqW5kiL8Pm9TDVwT1qG010_lSt_2Lv-L9nH_vZBWMy0 -O cudnn.tgz

其中cuda_9.2.148_396.37_linux为cuda文件，cuda_9.2.148.1_linux为cuda补丁，cudnn-9.2-linux-x64-v7.5.0.tgz为cudnn文件。

sh cuda_9.2.148_396.37_linux
sh cuda_9.2.148.1_linux
tar -zxvf cudnn.tgz
cp cuda/include/cudnn.h /usr/local/cuda/include
cp cuda/lib64/libcudnn* /usr/local/cuda/lib64
chmod a+r /usr/local/cuda/include/cudnn.h /usr/local/cuda/lib64/libcudnn*

使用nvcc --version查看cuda版本
使用cat /usr/local/cuda/include/cudnn.h | grep CUDNN_MAJOR -A 2

3. 安装Pytorch

conda install pytorch torchvision cudatoolkit=9.2 -c pytorch

如果安装比较慢，可以设成清华镜像，或者直接下载pytorch压缩包，conda install安装。

https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch/linux-64/pytorch-1.4.0-py3.6_cuda9.2.148_cudnn7.6.3_0.tar.bz2

实验代码，如果不出错，说明安装成功。

# Code in file tensor/two_layer_net_tensor.py
import torch
device = torch.device('cuda') # Uncomment this to run on GPU

# N is batch size; D_in is input dimension;
# H is hidden dimension; D_out is output dimension.
N, D_in, H, D_out = 64, 1000, 100, 10

# Create random input and output data
x = torch.randn(N, D_in, device=device)