Mintisan

On the way to be a practical idealism

首页 归档 标签 关于

Ubuntu 20.04 eGPU CUDA 开发环境搭建

2024-05-09 28 min read # 深度学习 # 开发环境

因为要用学习深度学习的剪枝量化，看到 AI Model Efficiency Toolkit (AIMET)
 最近更新还行，然后就去安装，发现只支持 Linux 系统，Ubuntu 的系统也有要求，然后加支持特定版本的 CUDA 以及特定版本的 Torch 和 Tensorflow，然后就有了这一篇记录。坑还是挺多的，记录一下方便后续搭建。

安装系统

安装完 Ubuntu 20.04 后重启，如果看到如下提示升级 22.04，不能升级，点击 Don't Upgrade。

并且点击 ok 确认不进行升级。

安装 no-machine

需要预先准备好离线 deb 安装包，并用 dbkg 进行本地安装如下【比较好的是安装完设置好了自动重启】：

mintisan@mintisan-NUC9VXQNX:~/UbuntuApps$ sudo dpkg -i ./nomachine_8.11.3_4_amd64.deb 
[sudo] password for mintisan: 
Selecting previously unselected package nomachine.
(Reading database ... 180009 files and directories currently installed.)
Preparing to unpack ./nomachine_8.11.3_4_amd64.deb ...
Unpacking nomachine (8.11.3-4) ...
Setting up nomachine (8.11.3-4) ...
NX> 700 Starting installation at: Sun, 19 May 2024 21:55:08.
NX> 700 Using installation profile: Ubuntu.
NX> 700 Installation log is: /usr/NX/var/log/nxinstall.log.
NX> 700 Installing nxrunner version: 8.11.3.
NX> 700 Installing nxplayer version: 8.11.3.
NX> 700 Installing nxnode version: 8.11.3.
NX> 700 Installing nxserver version: 8.11.3.
NX> 700 Installation completed at: Sun, 19 May 2024 21:55:18.
NX> 700 NoMachine was configured to run the following services:
NX> 700 NX service on port: 4000

然后在打开 NoMachine Server 可以看到本地的地址以及端口如下：

本地连接，并用安装时候的账号和密码登录即可。

注意在显示的时候需要适应本地窗口分辨率如下，否则会导致桌面显示不全。

安装 tailscale

如果不安装 tailscale ，则终端啥的下载很慢，如下所示：
【当然也可以换源，不过后面还是需要全局翻墙，早换早省心】

mintisan@mintisan-NUC9VXQNX:~/UbuntuApps$ sudo apt-get update -y
Hit:1 http://security.ubuntu.com/ubuntu focal-security InRelease
Err:2 http://cn.archive.ubuntu.com/ubuntu focal InRelease     
  Temporary failure resolving 'cn.archive.ubuntu.com'
Err:3 http://cn.archive.ubuntu.com/ubuntu focal-updates InRelease
  Temporary failure resolving 'cn.archive.ubuntu.com'
Err:4 http://cn.archive.ubuntu.com/ubuntu focal-backports InRelease
  Temporary failure resolving 'cn.archive.ubuntu.com'
Reading package lists... Done           
W: Failed to fetch http://cn.archive.ubuntu.com/ubuntu/dists/focal/InRelease  Temporary failure resolving 'cn.archive.ubuntu.com'
W: Failed to fetch http://cn.archive.ubuntu.com/ubuntu/dists/focal-updates/InRelease  Temporary failure resolving 'cn.archive.ubuntu.com'
W: Failed to fetch http://cn.archive.ubuntu.com/ubuntu/dists/focal-backports/InRelease  Temporary failure resolving 'cn.archive.ubuntu.com'
W: Some index files failed to download. They have been ignored, or old ones used instead.

安装 tailscale 需要登录 tailscale，而我的 tailscale 是用 Google 作为账号，那么就又到了一个鸡生蛋，蛋生鸡的时候。只能借助局域网内其他的代理来让浏览器可以登录 Google。设置本地代理如下：

注：我用的手机端的 SurfBoard 来进行代理，PC 端的 Clash 也可以，也就支持局域网代理中转的都是可以的。当然如果你有全局科学上网环境，也是可以的。

此时，就可以在 Chrome 端登录 tailscale 账号了。接下来，用命令行安装 tailscale 如下【需要预先安装 curl 等常用工具】：

mintisan@mintisan-NUC9VXQNX:~/UbuntuApps$ curl -fsSL https://tailscale.com/install.sh | sh

Command 'curl' not found, but can be installed with:

sudo snap install curl  # version 8.1.2, or
sudo apt  install curl  # version 7.68.0-1ubuntu2.20

See 'snap info curl' for additional versions.
mintisan@mintisan-NUC9VXQNX:~/UbuntuApps$ sudo apt-get install openssh-server tmux git zsh neofetch htop vim python3-pip wget curl hwinfo -y
...
mintisan@mintisan-NUC9VXQNX:~/UbuntuApps$ curl -fsSL https://tailscale.com/install.sh | sh
Installing Tailscale for ubuntu focal, using method apt
...
Setting up tailscale (1.66.3) ...
Created symlink /etc/systemd/system/multi-user.target.wants/tailscaled.service → /lib/systemd/system/tailscaled.service.
+ [ false = true ]
+ set +x
Installation complete! Log in to start using Tailscale by running:

sudo tailscale up
mintisan@mintisan-NUC9VXQNX:~/UbuntuApps$ sudo tailscale up


To authenticate, visit:

	https://login.tailscale.com/a/1aeaea1801138c

Success.
Some peers are advertising routes but --accept-routes is false
mintisan@mintisan-NUC9VXQNX:~/UbuntuApps$

输入 sudo tailscale up 再打开输出的链接，完成新设备到账号的绑定，绑定后就可以通过以下命令完成使用外部节点，完成所有的流量通过 VPN 来中专访问。

查看可以用作 exist-node 的 IP 地址：

mintisan@mintisan-NUC9VXQNX:~/UbuntuApps$ tailscale ip
100.97.189.1
fd7a:115c:a1e0::9a01:bd01
mintisan@mintisan-NUC9VXQNX:~/UbuntuApps$ tailscale status
100.97.189.1    ubuntu-nuc9vxqnx     fovwin@      linux   -
100.70.247.65   bandwagong-768.tail1456b.ts.net ran.thinker@ linux   idle; offers exit node
100.85.156.97   gl-mt3000            fovwin@      linux   -
100.71.139.27   mac-mini-m1          fovwin@      macOS   offline
100.122.227.134 mac-studio           fovwin@      macOS   -
100.74.183.119  macbook-air-m3       fovwin@      macOS   offline
100.88.126.81   macbook-pro-2018     fovwin@      macOS   offline
100.70.233.114  nuc13-extreme-980pro fovwin@      windows offline
100.75.92.65    oneplus-a6010        fovwin@      android offline
100.123.114.32  r2s-istoreos         fovwin@      linux   offline
100.122.151.42  r4s-istoreos         fovwin@      linux   -
100.95.124.152  racknerd-6e9267-2t   fovwin@      linux   idle; offers exit node
100.103.22.149  racknerd-f16b55-4t   fovwin@      linux   idle; offers exit node
100.78.129.159  raspberrypi5         fovwin@      linux   -
100.117.50.87   ser-pro-5800-archlinux fovwin@      linux   -
100.108.77.59   ser6-pro-vest-windows fovwin@      windows offline
100.80.43.50    tencent-ubuntu-2mbps fovwin@      linux   -
100.120.74.6    thinkpad-pibo        fovwin@      windows offline
100.110.45.44   vmware-ecg-server    fovwin@      linux   offline
100.113.135.2   vps-bandwagon-1000g  fovwin@      linux   idle; offers exit node
100.70.95.62    vps-bandwagon-500g   fovwin@      linux   idle; offers exit node
100.81.105.47   vps-bandwagon-768g   fovwin@      linux   idle; offers exit node
100.102.191.33  xiaomi-13            fovwin@      android offline
mintisan@mintisan-NUC9VXQNX:~/UbuntuApps$ sudo tailscale up --exit-node=100.113.135.2 --reset
Some peers are advertising routes but --accept-routes is false

再输入以上命令后， no-machine 会自动断开，此时需要在 no-machine 中重新设置 tailscale 的地址如下：

此时在终端就可以看到 ping 通 Google 了，完成了本地全局科学上网。

mintisan@mintisan-NUC9VXQNX:~/UbuntuApps$ ping google.com
PING google.com (172.217.25.174) 56(84) bytes of data.
64 bytes from sin01s16-in-f14.1e100.net (172.217.25.174): icmp_seq=1 ttl=118 time=201 ms
64 bytes from sin01s16-in-f14.1e100.net (172.217.25.174): icmp_seq=2 ttl=118 time=204 ms
^C
--- google.com ping statistics ---
2 packets transmitted, 2 received, 0% packet loss, time 1000ms
rtt min/avg/max/mdev = 200.994/202.248/203.503/1.254 ms

然后把网络代理也可以关掉了。

安装系统常用工具

安装 oh-my-zsh 如下，并设置 zsh 为默认 shell：

mintisan@mintisan-NUC9VXQNX:~/UbuntuApps$ sh -c "$(wget https://raw.githubusercontent.com/ohmyzsh/ohmyzsh/master/tools/install.sh -O -)"
Cloning Oh My Zsh...
remote: Enumerating objects: 1416, done.
remote: Counting objects: 100% (1416/1416), done.
remote: Compressing objects: 100% (1361/1361), done.
remote: Total 1416 (delta 47), reused 1091 (delta 25), pack-reused 0
Receiving objects: 100% (1416/1416), 3.21 MiB | 6.79 MiB/s, done.
Resolving deltas: 100% (47/47), done.
From https://github.com/ohmyzsh/ohmyzsh
 * [new branch]      master                   -> origin/master
 * [new branch]      update/plugins/wd/0.3.0  -> origin/update/plugins/wd/0.3.0
 * [new branch]      update/plugins/wd/v0.7.0 -> origin/update/plugins/wd/v0.7.0
Branch 'master' set up to track remote branch 'master' from 'origin'.
Already on 'master'
/home/mintisan/UbuntuApps

Looking for an existing zsh config...
Using the Oh My Zsh template file and adding it to /home/mintisan/.zshrc.

Time to change your default shell to zsh:
Do you want to change your default shell to zsh? [Y/n] Invalid choice. Shell change skipped.
         __                                     __   
  ____  / /_     ____ ___  __  __   ____  _____/ /_  
 / __ \/ __ \   / __ `__ \/ / / /  /_  / / ___/ __ \ 
/ /_/ / / / /  / / / / / / /_/ /    / /_(__  ) / / / 
\____/_/ /_/  /_/ /_/ /_/\__, /    /___/____/_/ /_/  
                        /____/                       ....is now installed!


Before you scream Oh My Zsh! look over the `.zshrc` file to select plugins, themes, and options.

• Follow us on Twitter: @ohmyzsh
• Join our Discord community: Discord server
• Get stickers, t-shirts, coffee mugs and more: Planet Argon Shop

➜  UbuntuApps chsh -s $(which zsh)

Password: 
➜  UbuntuApps

将 open-ssh server 加入开机自启动，方便后续通过 ssh 命令行访问本机：

➜  UbuntuApps sudo systemctl status ssh

● ssh.service - OpenBSD Secure Shell server
     Loaded: loaded (/lib/systemd/system/ssh.service; enabled; vendor preset: enabled)
     Active: active (running) since Sun 2024-05-19 22:19:07 CST; 21min ago
       Docs: man:sshd(8)
             man:sshd_config(5)
   Main PID: 18109 (sshd)
      Tasks: 1 (limit: 76667)
     Memory: 1.0M
     CGroup: /system.slice/ssh.service
             └─18109 sshd: /usr/sbin/sshd -D [listener] 0 of 10-100 startups

5月 19 22:19:07 mintisan-NUC9VXQNX systemd[1]: Starting OpenBSD Secure Shell server...
5月 19 22:19:07 mintisan-NUC9VXQNX sshd[18109]: Server listening on 0.0.0.0 port 22.
5月 19 22:19:07 mintisan-NUC9VXQNX sshd[18109]: Server listening on :: port 22.
5月 19 22:19:07 mintisan-NUC9VXQNX systemd[1]: Started OpenBSD Secure Shell server.
➜  UbuntuApps sudo systemctl enable ssh

Synchronizing state of ssh.service with SysV service script with /lib/systemd/systemd-sysv-install.
Executing: /lib/systemd/systemd-sysv-install enable ssh
➜  UbuntuApps sudo systemctl restart ssh

安装 Nvidia 驱动环境

此时没有接入 eGPU，输入如下查询显卡设备相关的命令，是没有输出的。

➜  UbuntuApps ls /usr/src | grep nvidia

➜  UbuntuApps hwinfo --gfxcard --short   
graphics card:                                                  
                       Intel UHD Graphics 630 (Mobile)

Primary display adapter: #34
➜  UbuntuApps 
➜  UbuntuApps lspci | grep -i nvidia

此时，输入 poweroff 可以进行关机：

➜  UbuntuApps sudo poweroff

重启后，USB 属于热插拔，但是 eGPU 实验表明属于半个热插拔，可以随着系统启动；也可以系统启动后再启动 eGPU ，再把 USB 接上电脑；但是如果已经处于连接的状态，再拔掉 eGPU ，此时系统会卡死，鼠标键盘都不能动，需要强制系统关机重启恢复。正常连接后再未安装驱动的情况下，可以看到如下情况：

➜  UbuntuApps ls /usr/src | grep nvidia
➜  UbuntuApps lspci | grep -i nvidia
3c:00.0 VGA compatible controller: NVIDIA Corporation GP102 [GeForce GTX 1080 Ti] (rev a1)
3c:00.1 Audio device: NVIDIA Corporation GP102 HDMI Audio Controller (rev a1)
➜  UbuntuApps hwinfo --gfxcard --short   
graphics card:                                                  
                       nVidia GP102 [GeForce GTX 1080 Ti]
                       Intel UHD Graphics 630 (Mobile)

Primary display adapter: #39

打开限制级驱动，可以看到会自动罗列 nVidia 的驱动列表，如果没有接入显卡，则会显示为空。如果显卡已经被系统识别，就可以看到系统会自动列出可以使用的驱动版本。

然后拔掉鼠标键盘，外接显示器【最好用 HDMI 诱骗器接上，也就是一定要让后主机看起来有显示器，不然即使连接也无法看到桌面】，插上 eGPU 【我用 NUC9 的下面那个雷电口，这里只是记录下】，并给 eGPU 电源打开。再打开主机的电源启动系统，等待 tailscale 上线。

安装驱动

➜  ~ ls /usr/src | grep nvidia

➜  ~ sudo ubuntu-drivers list
[sudo] password for mintisan: 
nvidia-driver-535, (kernel modules provided by linux-modules-nvidia-535-generic-hwe-20.04)
nvidia-driver-450-server, (kernel modules provided by linux-modules-nvidia-450-server-generic-hwe-20.04)
nvidia-driver-390, (kernel modules provided by linux-modules-nvidia-390-generic-hwe-20.04)
nvidia-driver-470, (kernel modules provided by linux-modules-nvidia-470-generic-hwe-20.04)
nvidia-driver-535-server, (kernel modules provided by linux-modules-nvidia-535-server-generic-hwe-20.04)
nvidia-driver-470-server, (kernel modules provided by linux-modules-nvidia-470-server-generic-hwe-20.04)
nvidia-driver-418-server, (kernel modules provided by linux-modules-nvidia-418-server-generic-hwe-20.04)
➜  ~ sudo apt install nvidia-driver-535 -y
...
➜  ~ 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.

在安装完驱动版本后，可以用 nvidia-smi 可执行程序，但是如上所示提示无法和显卡通信。打开后可以看到已经装上了，并且可以看到 535 的版本最高支持 12.2 CUDA 版本。

安装完需要重启 sudo reboot，然后就可以用命令行 nvidia-smi 看到驱动安装.

也可以用 nvitop 来看 gpu 的实时状态

➜  ~ pip3 install nvitop
➜  ~ python3 -m nvitop

安装 CUDA

在选项设置时，需要去掉 Driver，只安装 CUDA 相关的内容即可，显示如下：

➜  NVIDIA-DRIVER-CUDA-CUDNN sudo sh cuda_11.8.0_520.61.05_linux.run 

[sudo] password for mintisan: 
===========
= Summary =
===========

Driver:   Not Selected
Toolkit:  Installed in /usr/local/cuda-11.8/

Please make sure that
 -   PATH includes /usr/local/cuda-11.8/bin
 -   LD_LIBRARY_PATH includes /usr/local/cuda-11.8/lib64, or, add /usr/local/cuda-11.8/lib64 to /etc/ld.so.conf and run ldconfig as root

To uninstall the CUDA Toolkit, run cuda-uninstaller in /usr/local/cuda-11.8/bin
***WARNING: Incomplete installation! This installation did not install the CUDA Driver. A driver of version at least 520.00 is required for CUDA 11.8 functionality to work.
To install the driver using this installer, run the following command, replacing <CudaInstaller> with the name of this run file:
    sudo <CudaInstaller>.run --silent --driver

Logfile is /var/log/cuda-installer.log

安装完后，就将 cuda 的路径添加到 shell 的配置文件中：

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda-11.8/lib64
export PATH=$PATH:/usr/local/cuda-11.8/bin
export CUDA_HOME=$CUDA_HOME:/usr/local/cuda-11.8

添加后， source 重新加载一下，即可以看到安装的 cuda 版本

  NVIDIA-DRIVER-CUDA-CUDNN nvcc --version
zsh: command not found: nvcc
➜  NVIDIA-DRIVER-CUDA-CUDNN vim ~/.zshrc 
➜  NVIDIA-DRIVER-CUDA-CUDNN 
➜  NVIDIA-DRIVER-CUDA-CUDNN 
➜  NVIDIA-DRIVER-CUDA-CUDNN 
➜  NVIDIA-DRIVER-CUDA-CUDNN source ~/.zshrc 
➜  NVIDIA-DRIVER-CUDA-CUDNN nvcc --version 
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2022 NVIDIA Corporation
Built on Wed_Sep_21_10:33:58_PDT_2022
Cuda compilation tools, release 11.8, V11.8.89
Build cuda_11.8.r11.8/compiler.31833905_0

注：可以参考这篇文章来进行过个 cuda 版本的管理。

安装 CUDNN

目前我们安装的 cuda 是 11.8 版本，查看这篇文章可以看到对应需要的 cudnn 版本，我们这里 下载并安装安装 8.1.1 版本

➜  NVIDIA-DRIVER-CUDA-CUDNN sudo dpkg -i libcudnn8_8.1.1.33-1+cuda11.2_amd64.deb
Selecting previously unselected package libcudnn8.
(Reading database ... 195106 files and directories currently installed.)
Preparing to unpack libcudnn8_8.1.1.33-1+cuda11.2_amd64.deb ...
Unpacking libcudnn8 (8.1.1.33-1+cuda11.2) ...
Setting up libcudnn8 (8.1.1.33-1+cuda11.2) ...
Processing triggers for libc-bin (2.31-0ubuntu9.9) ...
➜  NVIDIA-DRIVER-CUDA-CUDNN sudo dpkg -i libcudnn8-dev_8.1.1.33-1+cuda11.2_amd64.deb
Selecting previously unselected package libcudnn8-dev.
(Reading database ... 195124 files and directories currently installed.)
Preparing to unpack libcudnn8-dev_8.1.1.33-1+cuda11.2_amd64.deb ...
Unpacking libcudnn8-dev (8.1.1.33-1+cuda11.2) ...
Setting up libcudnn8-dev (8.1.1.33-1+cuda11.2) ...
update-alternatives: using /usr/include/x86_64-linux-gnu/cudnn_v8.h to provide /usr/include/cudnn.h (libcudnn) in auto mode
➜  NVIDIA-DRIVER-CUDA-CUDNN sudo dpkg -i libcudnn8-samples_8.1.1.33-1+cuda11.2_amd64.deb

Selecting previously unselected package libcudnn8-samples.
(Reading database ... 195138 files and directories currently installed.)
Preparing to unpack libcudnn8-samples_8.1.1.33-1+cuda11.2_amd64.deb ...
Unpacking libcudnn8-samples (8.1.1.33-1+cuda11.2) ...
Setting up libcudnn8-samples (8.1.1.33-1+cuda11.2) ...
➜  NVIDIA-DRIVER-CUDA-CUDNN cat /usr/include/cudnn_version.h | grep CUDNN_MAJOR -A 2
#define CUDNN_MAJOR 8
#define CUDNN_MINOR 1
#define CUDNN_PATCHLEVEL 1
--
#define CUDNN_VERSION (CUDNN_MAJOR * 1000 + CUDNN_MINOR * 100 + CUDNN_PATCHLEVEL)

#endif /* CUDNN_VERSION_H */

安装 torch 和 tensorflow

安装 tenorflow 并确认是否可用

➜  ~ pip3 install tensorflow-gpu==2.10.0
Collecting tensorflow-gpu==2.10.0
...
  ~ python3 -c "import tensorflow as tf; print(tf.config.list_physical_devices('GPU'))"

2024-05-10 21:55:04.044654: I tensorflow/core/platform/cpu_feature_guard.cc:193] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations:  AVX2 FMA
To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.
2024-05-10 21:55:04.156741: E tensorflow/stream_executor/cuda/cuda_blas.cc:2981] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered
2024-05-10 21:55:04.661505: W tensorflow/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'libnvinfer.so.7'; dlerror: libnvinfer.so.7: cannot open shared object file: No such file or directory; LD_LIBRARY_PATH: :/usr/local/cuda-11.2/lib64
2024-05-10 21:55:04.661560: W tensorflow/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'libnvinfer_plugin.so.7'; dlerror: libnvinfer_plugin.so.7: cannot open shared object file: No such file or directory; LD_LIBRARY_PATH: :/usr/local/cuda-11.2/lib64
2024-05-10 21:55:04.661570: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Cannot dlopen some TensorRT libraries. If you would like to use Nvidia GPU with TensorRT, please make sure the missing libraries mentioned above are installed properly.
2024-05-10 21:55:05.084535: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:980] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2024-05-10 21:55:05.103192: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:980] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2024-05-10 21:55:05.103398: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:980] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
[PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]
➜  ~

安装 torch 并验证是否可以正确加载 gpu

➜  NVIDIA-DRIVER-CUDA-CUDNN pip3 install torch==1.13.1+cu116 torchvision==0.14.1+cu116 --extra-index-url https://download.pytorch.org/whl/cu116
Looking in indexes: https://pypi.org/simple, https://download.pytorch.org/whl/cu116
Collecting torch==1.13.1+cu116
  Downloading https://download.pytorch.org/whl/cu116/torch-1.13.1%2Bcu116-cp38-cp38-linux_x86_64.whl (1977.9 MB)
     |████████████████████████████████| 1977.9 MB 13 kB/s 
Collecting torchvision==0.14.1+cu116
  Downloading https://download.pytorch.org/whl/cu116/torchvision-0.14.1%2Bcu116-cp38-cp38-linux_x86_64.whl (24.2 MB)
     |████████████████████████████████| 24.2 MB 11.7 MB/s 
Requirement already satisfied: typing-extensions in /home/mintisan/.local/lib/python3.8/site-packages (from torch==1.13.1+cu116) (4.12.0)
Requirement already satisfied: pillow!=8.3.*,>=5.3.0 in /usr/lib/python3/dist-packages (from torchvision==0.14.1+cu116) (7.0.0)
Requirement already satisfied: requests in /usr/lib/python3/dist-packages (from torchvision==0.14.1+cu116) (2.22.0)
Requirement already satisfied: numpy in /home/mintisan/.local/lib/python3.8/site-packages (from torchvision==0.14.1+cu116) (1.24.4)
Installing collected packages: torch, torchvision
  WARNING: The scripts convert-caffe2-to-onnx, convert-onnx-to-caffe2 and torchrun are installed in '/home/mintisan/.local/bin' which is not on PATH.
  Consider adding this directory to PATH or, if you prefer to suppress this warning, use --no-warn-script-location.
Successfully installed torch-1.13.1+cu116 torchvision-0.14.1+cu116
➜  NVIDIA-DRIVER-CUDA-CUDNN python3 -c 'import torch; print(torch.cuda.is_available())'
True

这里我们下载一个简单的 mnist demo 来进行训练测试看下 gpu 被使用。

安装 AIMET

在这里终于可以回答为啥之前的 torch 用那个版本，以及需要对应的 cuda 版本。。。

就是因为需要能用 AIMET，你可以在这里找到对 torch 以及 cuda 的版本。

➜  UbuntuApps pip3 install ./aimet_torch-torch_gpu_1.31.0-cp38-cp38-linux_x86_64.whl

Processing ./aimet_torch-torch_gpu_1.31.0-cp38-cp38-linux_x86_64.whl
Collecting protobuf==3.20.2
。。。
ERROR: Could not find a version that satisfies the requirement torch==1.13.1+cu116 (from aimet-torch==torch-gpu-1.31.0) (from versions: 1.4.0, 1.5.0, 1.5.1, 1.6.0, 1.7.0, 1.7.1, 1.8.0, 1.8.1, 1.9.0, 1.9.1, 1.10.0, 1.10.1, 1.10.2, 1.11.0, 1.12.0, 1.12.1, 1.13.0, 1.13.1, 2.0.0, 2.0.1, 2.1.0, 2.1.1, 2.1.2, 2.2.0, 2.2.1, 2.2.2, 2.3.0)
ERROR: No matching distribution found for torch==1.13.1+cu116 (from aimet-torch==torch-gpu-1.31.0)

➜  UbuntuApps pip3 install torch==1.13.1+cu116 torchvision==0.14.1+cu116 --extra-index-url https://download.pytorch.org/whl/cu116
Looking in indexes: https://pypi.org/simple, https://download.pytorch.org/whl/cu116
Collecting torch==1.13.1+cu116
  Downloading https://download.pytorch.org/whl/cu116/torch-1.13.1%2Bcu116-cp38-cp38-linux_x86_64.whl (1977.9 MB)
     |████████████████████████████████| 1977.9 MB 13 kB/s 
Collecting torchvision==0.14.1+cu116
  Downloading https://download.pytorch.org/whl/cu116/torchvision-0.14.1%2Bcu116-cp38-cp38-linux_x86_64.whl (24.2 MB)
     |████████████████████████████████| 24.2 MB 5.5 MB/s 
Requirement already satisfied: typing-extensions in /home/mintisan/.local/lib/python3.8/site-packages (from torch==1.13.1+cu116) (4.11.0)
Requirement already satisfied: pillow!=8.3.*,>=5.3.0 in /usr/lib/python3/dist-packages (from torchvision==0.14.1+cu116) (7.0.0)
Requirement already satisfied: requests in /usr/lib/python3/dist-packages (from torchvision==0.14.1+cu116) (2.22.0)
Requirement already satisfied: numpy in /home/mintisan/.local/lib/python3.8/site-packages (from torchvision==0.14.1+cu116) (1.24.4)
ERROR: torchaudio 2.3.0 has requirement torch==2.3.0, but you'll have torch 1.13.1+cu116 which is incompatible.
Installing collected packages: torch, torchvision
  Attempting uninstall: torch
    Found existing installation: torch 1.13.1
    Uninstalling torch-1.13.1:
      Successfully uninstalled torch-1.13.1
  WARNING: The scripts convert-caffe2-to-onnx, convert-onnx-to-caffe2 and torchrun are installed in '/home/mintisan/.local/bin' which is not on PATH.
  Consider adding this directory to PATH or, if you prefer to suppress this warning, use --no-warn-script-location.
  Attempting uninstall: torchvision
    Found existing installation: torchvision 0.18.0
    Uninstalling torchvision-0.18.0:
      Successfully uninstalled torchvision-0.18.0
Successfully installed torch-1.13.1+cu116 torchvision-0.14.1+cu116
➜  UbuntuApps python3
Python 3.8.10 (default, Nov 22 2023, 10:22:35) 
[GCC 9.4.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> 
>>> from aimet_torch.quantsim import QuantizationSimModel
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/mintisan/.local/lib/python3.8/site-packages/aimet_torch/quantsim.py", line 53, in <module>
    import aimet_common.libpymo as libpymo
ImportError: liblapacke.so.3: cannot open shared object file: No such file or directory
>>>

是因为少了一个库，参考这里的说明按安装 liblapacke-dev 后再导入就没有问题了。

➜  UbuntuApps sudo apt-get install liblapacke-dev
。。。
Setting up liblapacke:amd64 (3.9.0-1build1) ...
Setting up libtmglib-dev:amd64 (3.9.0-1build1) ...
Setting up liblapacke-dev:amd64 (3.9.0-1build1) ...
Processing triggers for libc-bin (2.31-0ubuntu9.9) ...
➜  UbuntuApps python3                            
Python 3.8.10 (default, Nov 22 2023, 10:22:35) 
[GCC 9.4.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> from aimet_torch.quantsim import QuantizationSimModel
2024-05-25 20:52:34,923 - root - INFO - AIMET
>>>

这里记录下 torch 和 tensorflow 的版本。

➜  UbuntuApps pip list | grep torch
aimet-torch                  torch-gpu-1.31.0    
torch                        1.13.1+cu116        
torchvision                  0.14.1+cu116        
➜  UbuntuApps pip list | grep tensorflow
tensorflow-estimator         2.10.0              
tensorflow-gpu               2.10.0              
tensorflow-io-gcs-filesystem 0.34.0

参考链接

Powered by Gridea RSS