Skip to content

Latest commit

 

History

History
157 lines (106 loc) · 5.98 KB

cuda-installation.md

File metadata and controls

157 lines (106 loc) · 5.98 KB
Raw

A practical guide for CUDA reinstallation

This is a practical guide for CUDA reinstallation if you have conflict packages. It is worth noting that this tutorial is for a reinstallation for nvidia drivers, CUDA-toolkits and CuDnn.

All solutions, actions, operations here are validated by Xiaodan Xing at May 3rd, 2022. They might not work in a date after this date.

Update Cuda and Cudnn

Credit to Best practice for upgrading CUDA and cuDNN for tensorflow and Install CUDA 11.2, cuDNN 8.1.0, PyTorch v1.8.0 (or v1.9.0), and python 3.9 on RTX3090 for deep learning.

Install NVIDIA DRIVER

Step 1: Remove original Cuda files. Check /usr/local/ to see if there's any folder named "Cuda" left. For me there were still folders named CUDA or CUDA-xx.x, so I removed them. Please be cautious for this operation.

sudo apt-get --purge remove cuda
sudo apt-get autoremove
dpkg --list |grep "^rc" | cut -d " " -f 3 | xargs sudo dpkg --purge

Step 2: Then set up the Unbuntu by

sudo apt-get update
sudo apt-get upgrade -y
sudo apt-get install -y build-essential cmake unzip pkg-config
sudo apt-get install -y libxmu-dev libxi-dev libglu1-mesa libglu1-mesa-dev
sudo apt-get install -y libjpeg-dev libpng-dev libtiff-dev
sudo apt-get install -y libavcodec-dev libavformat-dev libswscale-dev libv4l-dev
sudo apt-get install -y libxvidcore-dev libx264-dev
sudo apt-get install -y libgtk-3-dev
sudo apt-get install -y libopenblas-dev libatlas-base-dev liblapack-dev gfortran
sudo apt-get install -y libhdf5-serial-dev graphviz
sudo apt-get install -y python3-dev python3-tk python-imaging-tk
sudo apt-get install -y linux-image-generic linux-image-extra-virtual
sudo apt-get install -y linux-source linux-headers-generic

Step 3: Add graphics PPA

sudo add-apt-repository ppa:graphics-drivers/ppa
sudo apt-get update

Step 4: Search available drivers

ubuntu-drivers devices

图片

Step 5: Install the driver with the best version. I choose driver-460 here because of the requirement of StyleGAN2-ada-PyTorch. Please select 510 if you are working on an RTX3090 device.

sudo apt-get install nvidia-driver-460

Step 6: Reboot your computer after installation. After rebooting, verify the driver installation by

nvidia-smi

Install CUDA-toolkit-v11.2

Step 1. Install CUDA-toolkit by

wget https://developer.download.nvidia.com/compute/cuda/11.2.2/local_installers/cuda_11.2.2_460.32.03_linux.run
sudo sh cuda_11.2.2_460.32.03_linux.run

Remember: DO NOT check the option of installing the driver!!!

In Yifan's blog, this line is written in bold and italic texts. I guess this is super important, but I do not know what will happen if the driver is installed simultaneously. To not install the driver, the installation page of CUDA should be like

图片

After the installation, you will see the result below. 图片

Step 2. Update environmental variable

Set environmental variables in the ~/.bashrc by

cd ~
vim ./.bashrc

Then paste the below lines into your bashrc file.

export PATH=/usr/local/cuda-11.2/bin${PATH:+:${PATH}}
export LD_LIBRARY_PATH=/usr/local/cuda-11.2/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
export CUDA_HOME=/usr/local/cuda

Remember to update your bashrc file by

source ./.bashrc

Install cuDnn-v8.1.0

Step 1. Download cuDnn-v8.1.0 from here

Step 2. Select a correct version of CuDnn. I also tried cuDNN v8.1.1 (Feburary 26th, 2021), for CUDA 11.0,11.1 and 11.2,cuDNN v8.2.0 (April 23rd, 2021), for CUDA 11.x, cuDNN v8.2.0 (April 23rd, 2021), for CUDA 11.x and cuDNN v8.2.1 (June 7th, 2021), for CUDA 11.x. They all worked perfectly for me and my RTX3090.

tar -zvxf cudnn-11.2-linux-x64-v8.1.0.77.tgz

Step 3. Set up environmental variables for CuDnn by adding the following at the end of ~/.bashrc.

export LD_LIBRARY_PATH=/home/xiaodan/cuda/lib64:$LD_LIBRARY_PATH

Step 4. To verify the installation, check nvidia-smi, and nvcc -V.

Update PyTorch or downgrade PyTorch

If the target is to use PyTorch in your device (i.e., you do not care which version of pytorch you will be using) please find a PyTorch-v1.11.0 downloading command here.

For the official PyTorch implementation of StyleGAN2, PyTorch-v1.7.x is required. Please find corresponding commands here.

# CUDA 11.1
conda install pytorch==1.8.0 torchvision==0.9.0 torchaudio==0.8.0 cudatoolkit=11.1 -c pytorch -c conda-forge

I always find the above conda command not working for me. Please use the pip wheel downloading command instead.

# CUDA 11.0
pip install torch==1.7.1+cu110 torchvision==0.8.2+cu110 torchaudio==0.7.2 -f https://download.pytorch.org/whl/torch_stable.html

Please do not worry if the version of cudatoolkit here does not match the CUDA version we just downloaded. It won't be a problem if you run codes from StyleGAN2.

Problems I encountered

  1. When using apt-get update,
The following signatures couldn't be verified because the public key is not available (The key number)

It is because the key is not added into your device. Just add the number into your keyserver by

apt-key adv --keyserver keyserver.ubuntu.com --recv-keys 'The key number'