TensorFlow with Nvidia Setup on Arch Linux

Published . Nil revisions.

Software Requirements

Check the latest software version requirements that are compatible with the particular TensorFlow version to be installed. These are listed under the Linux GPU tested build configurations section of the build from source web-page on the tensorflow.org website.

For example, TensorFlow 2.12, requires:

  • Python versions between 3.8 and 3.11,
  • CUDA Deep Neural Network library (cuDNN) version 8.6, and
  • CUDA compilation tools version 11.8.

Often the required versions will be different to those available through the official Arch Linux repositories, hence the processes for acquiring the specific versions are detailed in the following subsections.

Before progressing further, ensure the Nvidia drivers are up-to-date:

sudo pacman -S nvidia

Python

Check the version of Python available on the system:

python --version

If no Python installation is available, or the version available is not desirable and/or compatible with TensorFlow, see the Compiling Python from Source Code guide on how to create specific Python executable versions from source code.

cuDNN

Check if cuDNN is already installed, and if so, what version it is:

pacman -Qi cudnn

If cuDNN is not installed, or the version is not compatible with TensorFlow, check the version of cuDNN available in the official Arch Linux repositories: 

pacman -Si cudnn

If this version matches the requirement of TensorFlow, then install the package: 

sudo pacman -S cudnn

Else, follow the guide for installing previous versions of Arch Linux packages, adapting it for the cuDNN package.

CUDA

Repeat the process detailed in the cuDNN subsection prior to ensure a TensorFlow compatible version of CUDA is also installed.

If one is installing CUDA for the first time, then after installation, re-login to ensure that the CUDA binaries are added to the system PATH. To confirm this occurred, run: 

echo $PATH

Then check that the following directories are listed (among others):

/opt/cuda/bin:/opt/cuda/nsight_compute:/opt/cuda/nsight_systems/bin:/usr/lib/jvm/default/bin

Installation

Create a virtual environment using the latest version of Python that is compatible with TensorFlow (see Python Virtual Environments for more information on setting up virtual environments): 

<python-executable> -m venv <path-to-venv>

Activate the virtual environment:

source <path-to-venv>/bin/activate

Ensure pip and setuptools are up-to-date:

pip install --upgrade pip
pip install --upgrade setuptools

Install TensorFlow:

pip install tensorflow

Check if TensorFlow is able to use the GPU:

python -c "import tensorflow as tf; print(tf.config.list_physical_devices('GPU'))"