Full Installation¶

Virtual environment¶

It is preferable to install mendevi in a virtual environment. Please refer to the pyenv main page. It is possible to use python3-venv or conda as well.

Install pyenv¶

First install the python dependencies then install pyenv.

Linux

Debian Ubuntu Mint

sudo apt install libedit-dev libncurses5-dev build-essential libssl-dev zlib1g-dev libbz2-dev libreadline-dev libsqlite3-dev curl git libncursesw5-dev xz-utils tk-dev libxml2-dev libxmlsec1-dev libffi-dev liblzma-dev llvm-bolt libbolt-dev

RHEL CentOS Fedora

sudo dnf install openssl11-devel --allowerasing
sudo dnf install gcc make patch zlib-devel bzip2 bzip2-devel readline-devel sqlite sqlite-devel openssl-devel tk-devel libffi-devel xz-devel llvm-bolt

Arch Manjaro

yay -S ncurses5-compat-libs

curl https://pyenv.run | bash
CONTENT=$'export PYENV_ROOT="$HOME/.pyenv"\ncommand -v pyenv > /dev/null || export PATH="$PYENV_ROOT/bin:$PATH"\neval "$(pyenv init -)"\neval "$(pyenv virtualenv-init -)"'
grep -q "$CONTENT" ~/.bashrc || echo "$CONTENT" >> ~/.bashrc
source ~/.bashrc

macOS

You could install dependencies by using Homebrew.

brew install ncurses
brew install openssl readline sqlite3 xz zlib tcl-tk llvm
brew link llvm --force
brew install pyenv

Windows

Warning

Windows is crap, so be prepared for a tedious and buggy installation! You should forget Microchiotte-Windaube and go straight to Linux before you pull out all your hair!

Good luck with the official guide.

If it fails, this debug link may help you.

Configure pyenv¶

Create the virtual environment.

pyenv update
# --enable-experimental-jit=yes
# --enable-bolt
export PYTHON_CONFIGURE_OPTS='--enable-shared --enable-optimizations --with-lto=yes'
export PYTHON_CFLAGS='-march=native -mtune=native'
pyenv install -v 3.14
pyenv virtualenv 3.14 mendevi
pyenv activate mendevi
pip install uv

Dependencies¶

ffmpeg¶

Warning

If you want the latest version, please build it with: Build ffmpeg. And DO NOT INSTALL the precompiled ffmpeg package.

mendevi has hard dependency on the FFmpeg package. You should install it first, please refer to the FFmpeg download page.

Linux

Debian Ubuntu Mint

sudo apt install ffmpeg

RHEL CentOS Fedora

sudo dnf install ffmpeg

Arch Manjaro

sudo pacman -S ffmpeg

OpenSUSE

sudo zypper install ffmpeg

FreeBSD

sudo pkg install ffmpeg

macOS

You could install the FFmpeg by using Homebrew.

brew install ffmpeg pkg-config

Windows

Warning

I refuse to believe that you have gone through the previous steps. As you are stubborn and stupid, I must remind you that the game is not yet won! It is still not too late to listen to the voice of wisdom, which is telling you to get rid of this massive piece of Windows shit for ever.

It is important to configure your environment variable to hook ffmpeg to mendevi. You can try to follow this guide for example.

lshw¶

To get the full environment information, mendevi use tha system package lshw.

Linux

Debian Ubuntu Mint

sudo apt install lshw

RHEL CentOS Fedora

sudo dnf install lshw

Arch Manjaro

sudo pacman -S lshw

perf¶

To access the RAPL sensors.

Linux

Debian Ubuntu Mint

if ! sudo apt install -y linux-perf; then
    sudo apt install -y linux-tools-common linux-tools-generic linux-tools-$(uname -r)
fi

RHEL CentOS Fedora

sudo dnf install perf

Arch Manjaro

sudo pacman -S perf

transmission (optional)¶

To be able to automatically retrieve and share datasets and videos. For more informations, please have a look on the transmission webpage.

Linux

Debian Ubuntu Mint

sudo apt install --reinstall transmission-daemon

RHEL CentOS Fedora

sudo dnf install transmission-daemon

macOS

You could install transmission by using Homebrew.

brew install transmission transmission-cli

Ensure that transmission-daemon is installed correctly with the good permissions:

echo '======== CHECK AND FIX "transmission-daemon" INSTALLATION ========'

# find transmission-daemon user
trans_user=$(systemctl show -p User transmission-daemon.service | cut -d= -f2)

# create a common group with $USER and $trans_user
if getent group transshare > /dev/null; then
    echo 'ok - the group transshare already exists'
else
    echo "create a transshare group with $USER and $trans_user..."
    sudo groupadd transshare
    sudo usermod -aG transshare $USER
    sudo usermod -aG transshare $trans_user
    sudo systemctl restart transmission-daemon  # apply the effect
    echo '    done'
fi

# set the transshare group permissioin to ~/.cache/mendevi
mkdir -p ~/.cache/mendevi
if ! sudo -u $trans_user ls ~/.cache/mendevi/ &>/dev/null; then
    echo 'give the permission in ~/.cache/mendevi for transshare...'
    sudo chgrp -R transshare ~/.cache/mendevi/
    sudo chmod -R g+w ~/.cache/mendevi/
    sudo find ~/.cache/mendevi/ -type d -exec chmod 2775 {} \;
    sudo find ~/.cache/mendevi/ -type f -exec chmod ug+rw {} \;
    sudo chmod o+x ~/
    sudo chmod o+x ~/.cache/
    echo '    done (perhaps reboot required)'
else
    echo "ok - $trans_user can access to ~/.cache/mendevi"
fi

# make transmission-daemon working
if [ -n "$(transmission-remote -l 2>/dev/null)" ]; then
    echo 'ok - transmission-daemon seems to be reachable'
else
    echo 'attempted transmission repair...'
    sudo systemctl stop transmission-daemon
    sudo sed -i 's|Type=notify|Type=simple|g' /lib/systemd/system/transmission-daemon.service
    sudo sed -i 's|--log-level=error|--log-level=info|g' /lib/systemd/system/transmission-daemon.service
    sudo sed -i 's|"rpc-authentication-required": true|"rpc-authentication-required": false|g' /etc/transmission-daemon/settings.json
    # sudo sed -i "s|User=$trans_user|User=$USER|g" /lib/systemd/system/transmission-daemon.service
    sudo systemctl daemon-reload
    sudo systemctl start transmission-daemon
    echo "type 'systemctl status transmission-daemon' to check the status"
fi

# enable transmission-daemon
if [ "$(systemctl is-enabled transmission-daemon 2>/dev/null)" = "disabled" ]; then
    echo 'enable transmission-daemon service...'
    sudo rm -f /etc/systemd/system/multi-user.target.wants/transmission-daemon.service
    sudo systemctl enable transmission-daemon
    echo '    done'
else
    echo 'ok - transmission-daemon is enabled'
fi

# open graphical interface in a web browser
xdg-open http://localhost:9091

pyav (optional)¶

Although it is installed automatically, it is better to install av manually to avoid redundancy with ffmpeg. Please refer to the PyAv installation guide.

Linux

uv pip -v install av --no-binary av

FreeBSD

uv pip -v install av --no-binary av

macOS

On MacOS you may have issues with regards to Python expecting gcc but finding clang. Try to export the following before installation:

export ARCHFLAGS=-Wno-error=unused-command-line-argument-hard-error-in-future
uv pip -v install av --no-binary av

Windows

Warning

I see that you insist on using Windows, this step is the most critical, good luck!

On Windows you must build from source in order to indicate the location of ffmpeg, pleas follow the PyAv windows installation guide and good luck!

cuda (optional)¶

If you have a NVIDIA GPU, please install CUDA then follow the PyTorch installation guide. Without CUDA, the software is not able to use the GPU (CPU only).

Linux

Follow the official NVIDIA CUDA Installation Guide for Linux. Or the non official guide.

Debian Ubuntu Mint

Install nvidia driver.

Warning

Do not install “cuda-toolkit”, witch is conflicual with the official drivers.

Prerequisite: This block can be executed multiple times

sudo apt install linux-headers-$(uname -r)
sudo dkms autoinstall
sudo modprobe nvidia

if dpkg -l | grep -q nvidia-driver; then
    INSTALLED_NVIDIA_DRIVER_VERSION=$(dpkg-query -W -f='${Version}' nvidia-driver)
    echo "the nvidia driver $INSTALLED_NVIDIA_DRIVER_VERSION is installed and working"
else
    echo "no nvidia drivers are installed"

    # add nvidia driver depos
    if ! apt-cache show cuda-drivers >/dev/null 2>&1; then
        echo "try to add the official nvidia depos"
        wget https://developer.download.nvidia.com/compute/cuda/repos/debian13/x86_64/cuda-keyring_1.1-1_all.deb
        sudo dpkg -i cuda-keyring_1.1-1_all.deb
        sudo rm cuda-keyring_1.1-1_all.deb
        sudo apt update
    else
        echo "the official nvidia depos is available"
    fi

    # detect available versions
    AVAILABLE_NVIDIA_DRIVER_VERSION=$(sudo apt show cuda-drivers 2>/dev/null | awk '/Version:/ {print $2}' | head -n1 | cut -d'-' -f1)
    echo "NVIDIA detected dirver version: $AVAILABLE_NVIDIA_DRIVER_VERSION"
    DRIVER_BRANCH=${AVAILABLE_NVIDIA_DRIVER_VERSION%%.*}

    # install cuda
    # sudo apt remove --purge 'nvidia-*' cuda* && sudo apt update && sudo apt autoremove
    sudo apt install linux-headers-$(uname -r)  # or sudo apt install linux-headers-amd64
    sudo apt install cuda-drivers-${DRIVER_BRANCH}
    sudo dkms build -m nvidia -v ${AVAILABLE_NVIDIA_DRIVER_VERSION}
    sudo dkms install -m nvidia -v ${AVAILABLE_NVIDIA_DRIVER_VERSION}

    echo "to verify the installation, re-execute the code block"
fi

# wget https://us.download.nvidia.com/tesla/580.126.20/NVIDIA-Linux-x86_64-580.126.20.run
# chmod +x NVIDIA-Linux-x86_64-580.126.20.run
# ./NVIDIA-Linux-x86_64-580.126.20.run

Install cuda context (without the driver)

sudo apt install nvtop

# install nvcc
if command -v nvidia-smi &> /dev/null; then
    CUDA_VERSION=$(nvidia-smi --version | grep -oP 'CUDA Version\s*:\s*\K[0-9\.]+')
    echo "detected cuda version: $CUDA_VERSION"
    if ! command -v nvcc &> /dev/null; then
        CUDA_MAJOR=${CUDA_VERSION%%.*}
        CUDA_COMPILER_PACKAGE=$(apt-cache search "^cuda-compiler-${CUDA_MAJOR}-" | awk '{print $1}' | sort -V | tail -n1)
        echo "try to install $CUDA_COMPILER_PACKAGE (see 'apt search cuda-compiler')"
        sudo apt install $CUDA_COMPILER_PACKAGE

        CONTENT=$'export PATH=/usr/local/cuda/bin:$PATH\nexport LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH'
        grep -q "$CONTENT" ~/.bashrc || echo "$CONTENT" >> ~/.bashrc
        source ~/.bashrc
    else
        echo "nvcc is installed"
    fi
else
    echo "nvidia-smi is not installed, you have to (re)install the nvidia drivers"
    echo "sudo apt remove --purge 'nvidia-*' cuda* && sudo apt update && sudo apt autoremove"
    echo "then rerun the previous code block"
fi

FreeBSD

It is not well supproted, you can try to follow this bsd link.

macOS

Follow the NVIDIA CUDA Installation Guide for Mac OS X.

Windows

Warning

Windows is crap, so be prepared for a tedious and buggy installation! You should forget Microchiotte-Windaube and go straight to Linux before you pull out all your hair!

It is important to configure your environment variable to hook cuda to PyTorch. Good luck with the official guide.

amf (optional)¶

If you have an AMD GPU, please install ROC then follow the PyTorch installation guide. Without ROC, the software is not able to use the GPU (CPU only).

Linux

Debian Ubuntu Mint

Il faut ajouter https://github.com/CosmicFusion/debian-amdgpu-pro To use the free drivers of debian ati, execute the following code:

# lspci -nn | grep -E "VGA|Display|3D"  # to see if AMD gpu is available

sudo apt install firmware-linux-nonfree firmware-amd-graphics libgl1-mesa-dri libglx-mesa0 mesa-vulkan-drivers xserver-xorg-video-all

# lsmod | grep amdgpu  # if nothing apears -> probleme

To use the non-free drivers (required for video encoding but optional for torch), execute the following code. It comes from amdgpu-install package. And it doses the installation given in the AMD documentation.

# install amdgpu-install
if ! command -v amdgpu-install --help &> /dev/null; then
    BASE_URL='https://repo.radeon.com/amdgpu-install/latest/ubuntu/noble'
    latest_deb=$(curl -s "$BASE_URL/" | grep -oP 'amdgpu-install_.*?_all\.deb' | sort -V | tail -n 1)
    if [[ -z "$latest_deb" ]]; then
        echo "failed to find a .deb file in $BASE_URL"
    else
        echo "last package: $latest_deb"
        wget --progress=dot:giga "$BASE_URL/$latest_deb" -O "/tmp/$latest_deb"
        sudo apt install "/tmp/$latest_deb"
    fi
else
    echo 'the amdgpu-install package is installed'
fi

# install drivers
sudo amdgpu-install --usecase=amf --accept-eula

Installation¶

Basic Installation¶

To install a minimal version of mendevi, follow Simple Installation. Overwise, you can follow Building From Source (For devellopers) for a full installation.

Building From Source (For devellopers)¶

To install the lastest development version from GitLab source, clone mendevi using git and install it using uv pip:

git clone https://gitlab.inria.fr/rrichard/mendevi.git ~/mendevi_git
cd ~/mendevi_git
pyenv activate mendevi  # to be sure to be in a virtual env
uv pip install -e .[doc,test]
mendevi test  # to test if the installation is ok

You can now also compile documentation locally Build Documentation.

CLI auto-completion¶

To enable autocompletion in the command line interface, you can follow this guide, which involves executing the following code:

_MENDEVI_COMPLETE=bash_source mendevi > ~/.mendevi-complete.bash
CONTENT=$'. ~/.mendevi-complete.bash'
grep -q "$CONTENT" ~/.bashrc || echo "$CONTENT" >> ~/.bashrc
source ~/.bashrc

Build Documentation¶

You can compile the documentation yourself after Full Installation by simply typing mendevi doc or:

cd docs/ && make clean && make html && cd -
firefox docs/build/html/index.html

Testing¶

To run the unitary tests, just type mendevi test.