1. Install CUDA
1.1 Check if CUDA is installed
Shortcut key win+r, enter cmd, enter nvcc -V in the command line to view the version information
If it is already installed, the version information will be displayed, please skip to the second step
1.2 If it is not installed, please search and open the NVIDIA control panel in the lower left corner
Click System Information to enter the component page to view the version of CUDA
As shown in the figure, the CUDA version of the author's computer is 11.7, so choose to download CUDA ≤ 11.7 version
Enter the official website CUDA Toolkit Archive | NVIDIA Developer CUDA Toolkit Archive | NVIDIA Developer , download the corresponding version of CUDA CUDA Toolkit Archive | NVIDIA Developer
The version depends on the individual situation, do not pursue a higher version of CUDA than your own
After the download is complete, you will get an .exe installer, double-click to open it and go to the next step. It is recommended to install the default path . If you have any questions about this step, you can check this article:
Use the shortcut key win+r, enter cmd, and enter nvcc -V in the command line to check whether CUDA is installed successfully. If it is installed, the version information will be displayed
2. Install cuDNN
Enter the official website https://developer.nvidia.com/rdp/cudnn-download , first register and log in, and check your CUDA version number corresponding to your own cuDNN (the CUDA and cuDNN version comparison table used to be required, but now Nvidia only provides two versions cuDNN, corresponding to CUDA 12.x and CUDA 11.x respectively)
After downloading the compressed package, decompress it and change the file name to cudnn
Copy the cudnn file and paste it to the root directory of C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7
Finally, configure the environment variable of the system Path
These two sentences are actually the directory where the lib64 and bin files are stored in the cudnn folder just renamed, pay attention to the version number when copying and pasting
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\extras\CUPTI\lib64
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cudnn\bin
3. Install Pytorch
Enter the official website Start Locally | PyTorch , select the version of Pytorch according to the CUDA version installed in the first step
Through the shortcut key win+r, enter cmd, enter the following paragraph in the command line, download Pytorch
Note: The author tried many times, but the download failed. In addition, the method of other authors to download dependencies from the network disk by parsing the link address is outdated, so we adopt another universal method
We enter the link provided by this pip (according to the actual situation), which is the Pytorch dependent download page
We need to download torch, torchaudio, torchvision
3.1 Download torch
Click the torch link, search for the required .whl file according to the python version and CUDA version crtl+f search
cu117 means CUDA version 11.7 cp39 means python version 3.9
As you can see, there are two official versions of torch-1.13.0 and 1.13.1, choose one (recommended to choose the smaller one), and then choose win or linux according to the operating system
Because the author's computer has python 3.9 CUDA 11.7 win operating system, the following .whl files are downloaded
3.2 download torchaudio
Similarly, choose the appropriate version
3.3 Download torchvision
Similarly, choose the appropriate version
3.4 Installation files
Store the above files in a folder, enter cmd in the directory box, and enter the console
Enter the paragraph that pip did not download successfully before
pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu117
4.0 test
Enter the following codes in sequence from the console, if True appears, the installation is successful!
import torch
torch.__version__
torch.cuda.is_available()
Referenced articles (workaround and installation content sections are outdated):
(133 messages) torch.cuda.is_available() returns false - solution_Nefu_lyh's Blog-CSDN Blog