本文主要讲ubuntu18.04配置NVIDIA GPU环境， 并安装配置PyTorch。
先确认GPU型号
方式一:
$ lspci | grep -i nvidia
01:00.0 VGA compatible controller: NVIDIA Corporation Device 249d (rev a1)
01:00.1 Audio device: NVIDIA Corporation Device 228b (rev a1)
方式二:
$ ubuntu-drivers devices
== /sys/devices/pci0000:00/0000:00:01.0/0000:01:00.0 ==
modalias : pci:v000010DEd0000249Dsv00001D05sd00001147bc03sc00i00
vendor : NVIDIA Corporation
manual_install: True
driver : nvidia-driver-510 - distro non-free
driver : nvidia-driver-510-server - distro non-free
driver : nvidia-driver-515-server - distro non-free
driver : nvidia-driver-515 - distro non-free
driver : nvidia-driver-470-server - distro non-free
driver : nvidia-driver-470 - distro non-free recommended
driver : xserver-xorg-video-nouveau - distro free builtin
方式二中， driver : nvidia-driver-470 - distro non-free recommended便是系统推荐的的GPU 驱动版本。
安装驱动
只安装驱动
直接安装推荐的版本:
sudo apt install nvidia-driver-470
驱动 + cuda 安装
一般我们做深度学习使用nvidia gpu都需要安装cuda，安装cuda之前一定要确认依赖cuda的版本，比如当前最新版的PyTorch支持到的cuda版本为11.6, 而cuda最新版已经到11.6了，故要安装对应的版本。
我们根据英伟达官网链接 安装。
根据需要选择适合的cuda版本，点击链接进入安装， 此处我选择11.6。
nvidia-cuda-install-01.jpg
点击链接之后进入下一步，依次选择操作系统、架构、发行版（版本）、安装方式后，便给出的安装信息。
ubuntu安装方式栏有三个选项，三个我都试过，没特殊原因推荐使用runfile(local)方式，我在使用deb(local)和deb(network)安装的时候，会直接安装最新版的cuda，而我需要确定的一个版本11.6。
此处我选择"Linux-x86_64-Ubuntu-18.04-runfile(local)"
nvidia-cuda-install-02.jpg
复制安装命令依次执行:
$ wget https://developer.download.nvidia.com/compute/cuda/11.6.2/local_installers/cuda_11.6.2_510.47.03_linux.run
$ sudo sh cuda_11.6.2_510.47.03_linux.run
===========
= Summary =
===========
Driver: Installed
Toolkit: Installed in /usr/local/cuda-11.6/
Please make sure that
- PATH includes /usr/local/cuda-11.6/bin
- LD_LIBRARY_PATH includes /usr/local/cuda-11.6/lib64, or, add /usr/local/cuda-11.6/lib64 to /etc/ld.so.conf and run ldconfig as root
To uninstall the CUDA Toolkit, run cuda-uninstaller in /usr/local/cuda-11.6/bin
To uninstall the NVIDIA Driver, run nvidia-uninstall
Logfile is /var/log/cuda-installer.log
安装完成后根据提示配置环境变量
# cuda
export PATH="/usr/local/cuda-11.6/bin:$PATH"
export LD_LIBRARY_PATH="/usr/local/cuda-11.6/lib64:$LD_LIBRARY_PATH"
重启系统
使用以上两种方式安装都需要重启系统
$ sudo reboot
查看安装是否成功:
$ nvidia-smi
Sat Jul 9 07:04:15 2022
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 510.47.03 Driver Version: 510.47.03 CUDA Version: 11.6 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|===============================+======================+======================|
| 0 NVIDIA GeForce ... Off | 00000000:01:00.0 Off | N/A |
| N/A 52C P0 35W / N/A | 5MiB / 8192MiB | 0% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=============================================================================|
| 0 N/A N/A 1140 G /usr/lib/xorg/Xorg 4MiB |
+-----------------------------------------------------------------------------+
安装PyTorch
安装基础包
我一般使用conda管理python环境, 安装PyTorch之前先创建一个命名为dl-gpu的python环境:
$ conda create -y -n dl-gpu python=3.8
激活环境
$ conda activate dl-gpu
然后开始安装PyTorch,
首先进入官网安装链接， 依次选择PyTorch Build、Your OS、Package、Language、Compute Platform
pytorch-install-01.jpg
此处我选择的是"Stable(1.12.0)-Linux-Conda-Python-CUDA11.6"，得到执行的命令
$ conda install pytorch torchvision torchaudio cudatoolkit=11.6 -c pytorch -c conda-forge
注意点:
- 如果是conda安装，官网有一个提示
NOTE: 'conda-forge' channel is required for cudatoolkit 11.6, 需要先安装conda-forge, 安装可参考链接 - cuda版本与PyTorch支持的版本要尽量一致
验证安装
执行上述安装命令
$ conda install pytorch torchvision torchaudio cudatoolkit=11.6 -c pytorch -c conda-forge
验证是否可用:
(dl-gpu) ➜ ~ python -c "import torch;print(torch.cuda.device_count())"
1
这里输出则说明GPU配置成功了。
安装Jupyter
个人比较喜欢使用JupyterLab，所以本次安装JupyterLab, 安装细节见文档
使用conda安装
$ conda install -c conda-forge jupyterlab
使用pip安装
pip install jupyterlab
启动Jupyter并验证Torch和GPU
$ jupyter-lab
import torch
# 查看显卡数量
print('GPU count: ', torch.cuda.device_count())
# 获取第一个GPU
gpu1 = torch.device(f'cuda:0')
print('GPU one: ', gpu1)
# 创建使用GPU的张量X, Y
X = torch.rand(2, 3, device=gpu1)
Y = torch.rand(3, 4, device=gpu1)
# 分别打印X、Y、以及X和Y的乘积
X, Y, torch.mm(X, Y)
pytorch-gpu-py-example.jpg
其中device='cuda:0'表示当前张量数据存储在第0个GPU上。
常见问题
常见问题1： 重启系统后无法连接驱动了
$ nvidia-smi
NVIDIA-SMI has failed because it couldn't communicate with the NVIDIA driver. Make sure that the latest NVIDIA driver is installed and running.
大概率是更新Linux内核引起的，
- 先检查驱动和cuda版本
$ nvcc -V
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2022 NVIDIA Corporation
Built on Tue_Mar__8_18:18:20_PST_2022
Cuda compilation tools, release 11.6, V11.6.124
Build cuda_11.6.r11.6/compiler.31057947_0
说明驱动和和cuda都是存在的。
- 查看已经安装的驱动的版本信息
$ ls /usr/src|grep nvidia
nvidia-510.47.03
这里我的驱动版本是: nvidia-510.47.03
- 使用dkms工具安装驱动
$ sudo apt install dkms
$ sudo dkms install -m nvidia -v 510.47.03
...
nvidia-peermem.ko:
Running module version sanity check.
- Original module
- No original module exists within this kernel
- Installation
- Installing to /lib/modules/5.4.0-121-generic/updates/dkms/
depmod...
DKMS: install completed.
等待完成后再次输入nvidia-smi命令就发现驱动好了
$ nvidia-smi
Sat Jul 9 08:14:07 2022
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 510.47.03 Driver Version: 510.47.03 CUDA Version: 11.6 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|===============================+======================+======================|
| 0 NVIDIA GeForce ... Off | 00000000:01:00.0 Off | N/A |
| N/A 56C P0 34W / N/A | 0MiB / 8192MiB | 0% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=============================================================================|
| No running processes found |
+-----------------------------------------------------------------------------+
为了避免出现此问题，可关闭linux自动更新:
编辑文件/etc/apt/apt.conf.d/10periodic, 将其中的所有值改为0
