Skip to content

Latest commit

 

History

History
87 lines (60 loc) · 3.41 KB

CUDA_SETUP.md

File metadata and controls

87 lines (60 loc) · 3.41 KB
Raw

CUDA Installation

  • As mentioned in the README, one of the easiest methods to get the framework up and running is probably with Conda -- you can use either Miniconda or Anaconda.

  • Here's a quick method to install Miniconda on Linux:

  1. Download the Miniconda installer:

    wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh

  2. Make the installer executable:

    chmod +x Miniconda3-latest-Linux-x86_64.sh

  3. Run the installer:

    ./Miniconda3-latest-Linux-x86_64.sh

    Follow the prompts and choose the installation directory.

  4. Initialize conda (this adds conda to your shell’s startup):

    source ~/miniconda3/bin/activate
conda init

  5. Restart your terminal or run:

    source ~/.bashrc

  • After that, Miniconda should be ready to use.

  • I recommend installing mamba and using it instead of conda commands for way faster usability, like so:

    conda install mamba
mamba create yolov8_env  
mamba activate yolov8_env

  • You can then install the required packages i.e. like this:

    mamba install -y pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia
mamba install -c conda-forge ffmpeg

  • (Note: the pytorch-cuda version can be newer than the one mentioned above, I have tested the program in various setups up to CUDA 12.4 and CUDA 12.6 -- however, given the convoluted nature of the toolkit, newer version might mean new issues.)

  • If you're running multiple projects on the same machine, I highly recommend prioritizing the conda/mamba environment libraries instead of your regular LD_LIBRARY_PATH. You can check out your LD_LIBRARY_PATH with:

    echo $LD_LIBRARY_PATH

  • Note that the echo command should show your conda libraries first, i.e.: $HOME/miniconda3/envs/yolov8_env.

  • If your conda environment specific libraries aren't showing up at first, you can use the following commands to create an activation script that sets LD_LIBRARY_PATH to include only the Conda environment's lib directory, assuming you are using miniconda3:

    export ENV_NAME="yolov8_env"

# Create the activation directory if it doesn't exist
mkdir -p $HOME/miniconda3/envs/$ENV_NAME/etc/conda/activate.d

# Create the activation script to set LD_LIBRARY_PATH
echo 'export LD_LIBRARY_PATH=$HOME/miniconda3/envs/'$ENV_NAME'/lib' > $HOME/miniconda3/envs/$ENV_NAME/etc/conda/activate.d/env_vars.sh

# Make the script executable
chmod +x $HOME/miniconda3/envs/$ENV_NAME/etc/conda/activate.d/env_vars.sh

  • If you're still missing libraries such as cudnn, although not recommended in many use cases, you can also install CUDA related packages via conda/mamba by using Anaconda.org's nvidia channel, like so:

    mamba install -c nvidia cudatoolkit cudnn

  • However, keep in mind that you should only do these installs inside a very case-specific conda environment, like the one created for the project, and that multiple overlapping installations may introduce problems.

  • Also, some users and developers have reported that the precompiled builds offered by Nvidia's channel aren't necessarily always compiled with all the bells and whistles intact.

  • With all this said, good luck.