Install NVIDIA driver, CUDA, and CUDNN as described in here. Setup base and virtual python environment as in here. You can look at my sample setup functions below. These two function can be put in ~/.zshrc or ~/.zprofile (or ~/.bashrc and ~/.profile if you are using bash).
function setup_pyenv() {
export PYENV_ROOT="$HOME/.pyenv"
export PATH="$PYENV_ROOT/bin:$PATH"
if command -v pyenv 1>/dev/null 2>&1; then
eval "$(pyenv init -)"
fi
}
function setup_cudasupport_python() {
setup_pyenv
source $HOME/Pyenvs/cudasupport/bin/activate
}
You have 2 choices: using python-tensorflow-opt-cuda prebuilt and maintained by AUR, or build tensorflow and maintain (by maintaining, I just mean updating the source and rebuilding it when needed, not like writing fixes and patches) a tensorflow package locally, on your own. A caveat of the first choice is that you will have to use the native python, which I always highly recommend people not to touch. If you want to do it anyway, you can easily install the package with yay.
yay -S python-tensorflow-opt-cuda
If you want to build tensorflow from the source. Follow through the next steps. Notice that, from here, we are following the official tensorflow website’s instructions.
Now you have a virtual python environment called cudasupport, go ahead and set it up (open a new terminal, locate and run the activate script corresponding to the environment, or run setup_cuda_support if you chose to use my setup functions mentioned earlier).
We will need to upgrade pip and wheel to the latest version, install numpy without --upgrade and keras_preprocessing with --no-deps. Notice: if you are not using virtual environment, adding --user option.
pip install --upgrade pip setuptools wheel
pip install numpy
pip install keras_preprocessing --no-deps
Getting the source code. I make a folder in my virtual environment base directory (same level as bin, lib and include) and name it tensorflow-build. This is the folder that I will use to keep the tensorflow’s source code.
cd ./tensorflow-build # Notice that my current folder is virtual environment base dir
git clone https://github.com/tensorflow/tensorflow.git
cd tensorflow
git checkout r2.4
I choose the r2.4 release here. In my case, I can successfully build it with gcc (7.5.0), CUDA (11.1.105), CUDNN (8.0.5), TensorRT (7.0) (you will not need TensorRT for your laptop or personal workstation, I just want to see if my build can work with TensorRT 7.0), bazel (3.1.0), base python (3.7.3), and numpy (1.19.4). In the first try to build it with gcc (10.2.0) which is the default version on Arch release (2020.11.01), I encounter a lot of issues. I will post them here later. So, please install gcc-7 with yay before you proceed.
To be continued…