🛤️Install Pytorch on a GPU Server

This is Harry, as I used to install Pytorch on WSL for specific need, in this note, I take install torch==1.10.1+cu113 for example.

Instructions

Install CUDA

First

In Winows cmd, use nvidia-smi to check the version of NVIDIA drive. For example, in the figure below, the version is 522.25

If you haven't installed a NVIDIA driver on Windows, go to NVIDIA Driver Downloads page to search, download and install the lastest version of NVIDIA driver that you can install for Winows due to your hardware and software conditions.

After installation, excute nvidia-smi again to check.

Second

Note: the new Windows NVIDIA drivers have built-in support for WSL 2, which means that you only need to have the NVIDIA drivers in Windows, not in WSL 2. You only need to install the CUDA Toolkit for WSL 2.

In the figure above, we konw that the latest version of CUDA driver that the NVIDIA driver support is 11.8 from the return of nvidia-smi. In this note, I download CUDA 11.3.0 for WSL-Ubuntu.

1. Find Correct Version: To install CUDA driver for WSL-Ubuntu, go to CUDA Toolkit Archive page, choose your target CUDA driver and follow the official instructions. For information about debugging, go to Additions.

   

   CUDA download steps
2. Install: If you use runfile (local) to install CUDA, enter accept for user license agreement and customize installation components (CUDA Samples, Demo Suite and Documentation are not needed). After customization, enter install and it will start to install.

   

   CUDA user license agreement

   

   customize installation components

   

   Successful installation
3. Configure Environment Variables: enter sudo vim ~/.bashrc and append the following code to the end:

   export PATH=/usr/local/cuda-11.3/bin${PATH:+:${PATH}}
   export LD_LIBRARY_PATH=/usr/local/cuda-11.3/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}

   Then, if nothing goes wrong, enter nvcc -V and you can view the version of CUDA driver on WSL-Ubuntu (11.3), which is different from the version on my host Windows (11.8)

   

   nvcc --version

Install Pytorch

Go to Previous Pytorch Version page to get official instructions to download and install torch, which is corresponding to your CUDA driver and system.

In this case, I use:

pip install torch==1.10.1+cu113 torchvision==0.11.2+cu113 torchaudio==0.10.1 -f https://download.pytorch.org/whl/cu113/torch_stable.html

where 1.10.1 refers to the version of torch and cu113 refers to the lowest version of CUDA that the torch supports.

After installation, use the following code to check whether pytorch is available on your GPU server:

import torch
print(torch.cuda.is_available())

The return True is expected. If False is returned, check your installation steps!

Additions

import torch
print(torch.__version__)  # view torch version
print(torch.cuda.is_available()) # view whether CUDA is available
print(torch.cuda.device_count()) # view number of available GPU
print(torch.version.cuda) # view CUDA version

Additions

ERROR: Missing gcc. gcc is required to continue.

When install CUDA following official instructions with runfile (local), the following error may occur:

Failed to verify gcc version. See log at /var/log/cuda-installer.log for details.

In /var/log/cuda-installer.log, you may see [error] Missing gcc. gcc is required to continue. Follow the code below and you can solve it:

sudo apt update
sudo apt install build-essential

use gcc --version to verify whether gcc has been successfully install. The default version of GCC available in the Ubuntu 20.04 is 9.3.0:

gcc (Ubuntu 9.3.0-10ubuntu2) 9.3.0
Copyright (C) 2019 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

References

1. WSL安装Ubuntu及配置Pytorch

2. WSL 2 Ubuntu 20.04 配置CUDA 原创

3. 以Runfile的方式安装CUDA 11