This set of environment is really too complex. There are too many pits. After more than half a year, it finally succeeded today.

Graphics card driver

Direct installation system recommends graphics card driver, with the lowest error probability.

sudo ubuntu-drivers autoinstall

View installation status

nvidia-smi

The graphics card version and CUDA support will appear

CUDA

First of all, thank you for your original post:
https://blog.csdn.net/weixin_43491255/article/details/118549032
https://blog.csdn.net/a563562675/article/details/119458550
Disable nouveau first

sudo gedit /etc/modprobe.d/blacklist.conf 

Add these two lines to save

blacklist nouveau
options nouveau modeset=0

Refresh

 sudo update-initramfs -u

Restart the computer, be sure to restart. Then enter this command. If nothing is output, it is successful.

 lsmod | grep nouveau

Install CUDA

First, download the CUDA installation package (runfile format) you need here
Because you know the reason, it is recommended to open it with a kick, otherwise you can't get in.
https://developer.nvidia.com/cuda-toolkit-archive
Be sure to pay attention to the version!!!
I recommend CUDA 11.3.0 + cudnn 8.2.1 + Python 1.10
The combination of the three is very troublesome. I tested it successfully in person, and I can't guarantee other combinations.
Because you know the reason, wget downloads very slowly. Here is a trick. Copy the link behind wget, paste it to the browser that can use kicks, and download it directly with the browser. I was about 3M/s, and the speed was pretty good. The file is large, 2.8G, Download patiently.
Enter this installation after the installation:

sudo sh cuda_11.3.0_465.19.01_linux.run

This will appear next. Press enter in the Driver, do not install the Driver (we have installed the Driver just now), and then select Install to install.
After installation, open the environment variable

vim ~/.bashrc

Add these two lines to save. Note that this is 11.3. If you are in another version, you should change it to your own version

export PATH="/usr/local/cuda-11.3/bin:$PATH"
export LD_LIBRARY_PATH="/usr/lcoal/cuda-11.3/lib64:$LD_LIBRARY_PATH"

Refresh environment variables

source ~/.bashrc

test

Terminal input

cd /usr/local/cuda-11.3/samples/1_Utilities/deviceQuery

sudo make

./deviceQuery

If so, it indicates success, and the available equipment information of CUDA is displayed.

CUDNN

install

Download three deb packages first. Be sure to use deb. The probability of error is small
https://developer.nvidia.com/rdp/cudnn-archive
Mine is 11.3. Select 8.2.1 for 11.x

Finally, install the three packages

sudo dpkg -i xxxxxxx.deb

test

Just now, a package is cudnn's own sample. Run the test

By the way, run this before testing to prevent errors later

sudo apt-get install libfreeimage3 libfreeimage-dev

Start running sample

cd /usr/src/cudnn_samples_v7

cp -r /usr/src/cudnn_samples_v7/ $HOME

cd $HOME/cudnn_samples_v7/mnistCUDNN

make clean && make

./mnistCUDNN

If Test passed appears, the installation is successful

Check out the cudnn version

find / -name cudnn_version.h 2>&1 | grep -v "Permission denied"

cat /usr/include/cudnn_version.h | grep CUDNN_MAJOR -A 2

Anaconda

Download the installation package from Tsinghua image source
https://mirrors.tuna.tsinghua.edu.cn/anaconda/archive/
After downloading, enter the command to install. Remember to select yes all the way

bash Anaconda3-2021.05-Linux-x86_64.sh

Open condarc to switch the source of conda

sudo gedit ~/.condarc

Replace the contents with this. The method of modifying defaults and the homepage of TUNA official website has been included.

channels:
  - defaults
show_channel_urls: true
default_channels:
  - https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
  - https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/r
  - https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/msys2
custom_channels:
  conda-forge: https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud
  msys2: https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud
  bioconda: https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud
  menpo: https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud
  pytorch: https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud
  simpleitk: https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud
report_errors: false
auto_activate_base: false

Finally, create a virtual environment to install pytorch

conda create -n alientorch((your own name)

Activate environment

activate alientorch

Install a wave of dependencies and test the source just now. If the speed is very fast, it will be great, and the pytoch will be very smooth in a while

conda install numpy mkl cffi

Pytorch

install

Enter the official website to view the installation commands corresponding to your version

https://pytorch.org/get-started/locally/
Enter the command to install pytorch in the virtual environment just now

conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch

The installation file is very large. Just wait patiently. If the source switch is successful, the download here will be fast.

test

Enter the virtual environment, enter the command test, and you're done!

python

import torch

torch.cuda.is_available()