背景
最近在学习PaddlePaddle
在各个显卡驱动版本的安装和使用,所以同时也学习如何在Ubuntu
安装和卸载CUDA和CUDNN
,在学习过程中,顺便记录学习过程。在供大家学习的同时,也在加强自己的记忆。本文章以卸载CUDA 8.0
和 CUDNN 7.05
为例,以安装CUDA 10.0 和 CUDNN 7.4.2
为例。
安装显卡驱动
禁用nouveau
驱动
sudo vim /etc/modprobe.d/blacklist.conf
在文本最后添加:
blacklist nouveau
options nouveau modeset=0
然后执行:
sudo update-initramfs -u
重启后,执行以下命令,如果没有屏幕输出,说明禁用nouveau
成功:
lsmod | grep nouveau
下载驱动
官网下载地址:https://www.nvidia.cn/Download/index.aspx?lang=cn ,根据自己显卡的情况下载对应版本的显卡驱动,比如显卡是RTX2070
:
下载完成之后会得到一个安装包,不同版本文件名可能不一样:
NVIDIA-Linux-x86_64-410.93.run
卸载旧驱动
以下操作都需要在命令界面操作,执行以下快捷键进入命令界面,并登录:
Ctrl-Alt+F1
执行以下命令禁用X-Window
服务,否则无法安装显卡驱动:
sudo service lightdm stop
执行以下三条命令卸载原有显卡驱动:
sudo apt-get remove --purge nvidia*
sudo chmod +x NVIDIA-Linux-x86_64-410.93.run
sudo ./NVIDIA-Linux-x86_64-410.93.run --uninstall
安装新驱动
直接执行驱动文件即可安装新驱动,一直默认即可:
sudo ./NVIDIA-Linux-x86_64-410.93.run
执行以下命令启动X-Window服务
sudo service lightdm start
最后执行重启命令,重启系统即可:
reboot
注意: 如果系统重启之后出现重复登录的情况,多数情况下都是安装了错误版本的显卡驱动。需要下载对应本身机器安装的显卡版本。
卸载CUDA
为什么一开始我就要卸载CUDA呢,这是因为换了显卡RTX2070,原本就安装了CUDA 8.0 和 CUDNN 7.0.5不能够正常使用,需要安装CUDA 10.0 和 CUDNN 7.4.2,所以要先卸载原来的CUDA。注意以下的命令都是在root用户
下操作的。
卸载CUDA很简单,一条命令就可以了,主要执行的是CUDA自带的卸载脚本,读者要根据自己的cuda版本找到卸载脚本:
sudo /usr/local/cuda-8.0/bin/uninstall_cuda_8.0.pl
卸载之后,还有一些残留的文件夹,之前安装的是CUDA 8.0。可以一并删除:
sudo rm -rf /usr/local/cuda-8.0/
这样就算卸载完了CUDA。
安装CUDA
安装的CUDA和CUDNN版本:
- CUDA 10.0
- CUDNN 7.4.2
接下来的安装步骤都是在root用户下操作的。
下载和安装CUDA
我们可以在官网:CUDA10下载页面,
下载符合自己系统版本的CUDA。页面如下:
下载完成之后,给文件赋予执行权限:
chmod +x cuda_10.0.130_410.48_linux.run
执行安装包,开始安装:
./cuda_10.0.130_410.48_linux.run
开始安装之后,需要阅读说明,可以使用Ctrl + C
直接阅读完成,或者使用空格键慢慢阅读。然后进行配置,我这里说明一下:
(是否同意条款,必须同意才能继续安装)
accept/decline/quit: accept(这里不要安装驱动,因为已经安装最新的驱动了,否则可能会安装旧版本的显卡驱动,导致重复登录的情况)
Install NVIDIA Accelerated Graphics Driver for Linux-x86_64 410.48?
(y)es/(n)o/(q)uit: nInstall the CUDA 10.0 Toolkit?(是否安装CUDA 10 ,这里必须要安装)
(y)es/(n)o/(q)uit: yEnter Toolkit Location(安装路径,使用默认,直接回车就行)[ default is /usr/local/cuda-10.0 ]: Do you want to install a symbolic link at /usr/local/cuda?(同意创建软链接)
(y)es/(n)o/(q)uit: yInstall the CUDA 10.0 Samples?(不用安装测试,本身就有了)
(y)es/(n)o/(q)uit: nInstalling the CUDA Toolkit in /usr/local/cuda-10.0 ...(开始安装)
安装完成之后,可以配置他们的环境变量,在vim ~/.bashrc
的最后加上以下配置信息:
export CUDA_HOME=/usr/local/cuda-10.0
export LD_LIBRARY_PATH=${CUDA_HOME}/lib64
export PATH=${CUDA_HOME}/bin:${PATH}
最后使用命令source ~/.bashrc
使它生效。
可以使用命令nvcc -V
查看安装的版本信息:
test@test:~$ nvcc -V
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2018 NVIDIA Corporation
Built on Sat_Aug_25_21:08:01_CDT_2018
Cuda compilation tools, release 10.0, V10.0.130
测试安装是否成功
执行以下几条命令:
cd /usr/local/cuda-10.0/samples/1_Utilities/deviceQuery
make
./deviceQuery
正常情况下输出:
./deviceQuery Starting...CUDA Device Query (Runtime API) version (CUDART static linking)Detected 1 CUDA Capable device(s)Device 0: "GeForce RTX 2070"CUDA Driver Version / Runtime Version 10.0 / 10.0CUDA Capability Major/Minor version number: 7.5Total amount of global memory: 7950 MBytes (8335982592 bytes)(36) Multiprocessors, ( 64) CUDA Cores/MP: 2304 CUDA CoresGPU Max Clock rate: 1620 MHz (1.62 GHz)Memory Clock rate: 7001 MhzMemory Bus Width: 256-bitL2 Cache Size: 4194304 bytesMaximum Texture Dimension Size (x,y,z) 1D=(131072), 2D=(131072, 65536), 3D=(16384, 16384, 16384)Maximum Layered 1D Texture Size, (num) layers 1D=(32768), 2048 layersMaximum Layered 2D Texture Size, (num) layers 2D=(32768, 32768), 2048 layersTotal amount of constant memory: 65536 bytesTotal amount of shared memory per block: 49152 bytesTotal number of registers available per block: 65536Warp size: 32Maximum number of threads per multiprocessor: 1024Maximum number of threads per block: 1024Max dimension size of a thread block (x,y,z): (1024, 1024, 64)Max dimension size of a grid size (x,y,z): (2147483647, 65535, 65535)Maximum memory pitch: 2147483647 bytesTexture alignment: 512 bytesConcurrent copy and kernel execution: Yes with 3 copy engine(s)Run time limit on kernels: YesIntegrated GPU sharing Host Memory: NoSupport host page-locked memory mapping: YesAlignment requirement for Surfaces: YesDevice has ECC support: DisabledDevice supports Unified Addressing (UVA): YesDevice supports Compute Preemption: YesSupports Cooperative Kernel Launch: YesSupports MultiDevice Co-op Kernel Launch: YesDevice PCI Domain ID / Bus ID / location ID: 0 / 1 / 0Compute Mode:< Default (multiple host threads can use ::cudaSetDevice() with device simultaneously) >deviceQuery, CUDA Driver = CUDART, CUDA Driver Version = 10.0, CUDA Runtime Version = 10.0, NumDevs = 1
Result = PASS
下载和安装CUDNN
进入到CUDNN的下载官网:https://developer.nvidia.com/rdp/cudnn-download ,然点击Download开始选择下载版本,当然在下载之前还有登录,选择版本界面如下,我们选择cuDNN Library for Linux
:\
下载之后是一个压缩包,如下:
cudnn-10.0-linux-x64-v7.4.2.24.tgz
然后对它进行解压,命令如下:
tar -zxvf cudnn-10.0-linux-x64-v7.4.2.24.tgz
解压之后可以得到以下文件:
cuda/include/cudnn.h
cuda/NVIDIA_SLA_cuDNN_Support.txt
cuda/lib64/libcudnn.so
cuda/lib64/libcudnn.so.7
cuda/lib64/libcudnn.so.7.4.2
cuda/lib64/libcudnn_static.a
使用以下两条命令复制这些文件到CUDA目录下:
cp cuda/lib64/* /usr/local/cuda-10.0/lib64/
cp cuda/include/* /usr/local/cuda-10.0/include/
拷贝完成之后,可以使用以下命令查看CUDNN的版本信息:
cat /usr/local/cuda/include/cudnn.h | grep CUDNN_MAJOR -A 2
测试安装结果
到这里就已经完成了CUDA 10 和 CUDNN 7.4.2
的安装。可以安装对应的Pytorch的GPU版本测试是否可以正常使用了。安装如下:
pip3 install https://download.pytorch.org/whl/cu100/torch-1.0.0-cp35-cp35m-linux_x86_64.whl
pip3 install torchvision
然后使用以下的程序测试安装情况:
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import torch.backends.cudnn as cudnn
from torchvision import datasets, transformsclass Net(nn.Module):def __init__(self):super(Net, self).__init__()self.conv1 = nn.Conv2d(1, 10, kernel_size=5)self.conv2 = nn.Conv2d(10, 20, kernel_size=5)self.conv2_drop = nn.Dropout2d()self.fc1 = nn.Linear(320, 50)self.fc2 = nn.Linear(50, 10)def forward(self, x):x = F.relu(F.max_pool2d(self.conv1(x), 2))x = F.relu(F.max_pool2d(self.conv2_drop(self.conv2(x)), 2))x = x.view(-1, 320)x = F.relu(self.fc1(x))x = F.dropout(x, training=self.training)x = self.fc2(x)return F.log_softmax(x, dim=1)def train(model, device, train_loader, optimizer, epoch):model.train()for batch_idx, (data, target) in enumerate(train_loader):data, target = data.to(device), target.to(device)optimizer.zero_grad()output = model(data)loss = F.nll_loss(output, target)loss.backward()optimizer.step()if batch_idx % 10 == 0:print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(epoch, batch_idx * len(data), len(train_loader.dataset),100. * batch_idx / len(train_loader), loss.item()))def main():cudnn.benchmark = Truetorch.manual_seed(1)device = torch.device("cuda")kwargs = {'num_workers': 1, 'pin_memory': True}train_loader = torch.utils.data.DataLoader(datasets.MNIST('../data', train=True, download=True,transform=transforms.Compose([transforms.ToTensor(),transforms.Normalize((0.1307,), (0.3081,))])),batch_size=64, shuffle=True, **kwargs)model = Net().to(device)optimizer = optim.SGD(model.parameters(), lr=0.01, momentum=0.5)for epoch in range(1, 11):train(model, device, train_loader, optimizer, epoch)if __name__ == '__main__':main()
如果正常输出一下以下信息,证明已经安装成了:
Train Epoch: 1 [0/60000 (0%)] Loss: 2.365850
Train Epoch: 1 [640/60000 (1%)] Loss: 2.305295
Train Epoch: 1 [1280/60000 (2%)] Loss: 2.301407
Train Epoch: 1 [1920/60000 (3%)] Loss: 2.316538
Train Epoch: 1 [2560/60000 (4%)] Loss: 2.255809
Train Epoch: 1 [3200/60000 (5%)] Loss: 2.224511
Train Epoch: 1 [3840/60000 (6%)] Loss: 2.216569
Train Epoch: 1 [4480/60000 (7%)] Loss: 2.181396
参考资料
https://developer.nvidia.com
https://www.cnblogs.com/luofeel/p/8654964.html