Install tensorflow-gpu on Fedora 29

I want to run the notebook from Guillaume Lambard on our server. Operating system is a Fedora 29 because it is the last version supported by the NVIDIA cuda toolkit. The GPU is a TESLA K-20.

Install the CUDA tookit

dnf config-manager --add-repo http://developer.download.nvidia.com/compute/cuda/repos/fedora$(rpm -E %fedora)/x86_64/cuda-fedora$(rpm -E %fedora).repo
dnf clean all
dnf install gcc-c++ mesa-libGLU-devel libX11-devel libXi-devel libXmu-devel
dnf install kernel-devel-$(uname -r) kernel-headers-$(uname -r)
sudo dnf clean expire-cache
sudo dnf install cuda cuda-drivers cuda-toolkit-10.2

You need to reboot the server to load the driver.

Check your GPU with

nvidia-smi

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 440.33.01    Driver Version: 440.33.01    CUDA Version: 10.2     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  Tesla K20m          Off  | 00000000:04:00.0 Off |                    0 |
| N/A   31C    P0    54W / 225W |      0MiB /  4743MiB |     99%      Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID   Type   Process name                             Usage      |
|=============================================================================|
|  No running processes found                                                 |
+-----------------------------------------------------------------------------+

Install cuDNN

cuDDN is available with a NVIDIA developer free account. I take the CENTOS/RedHat versions for CUDA 10.1:

  • libcudnn7-7.6.5.33-1.cuda10.1.x86_64.rpm
  • libcudnn7-devel-7.6.5.33-1.cuda10.1.x86_64.rpm
  • libcudnn7-doc-7.6.5.33-1.cuda10.1.x86_64.rpm
rpm -Uvh libcuddn*.rpm

Install tensorflow-gpu

The tensorflow-gpu pip package doesn’t work because it is built with CUDA 10.0. And CUDA 10.0 is not available with Fedora 29. So you need to build it from scratch. To build tensorflow you need some dependencies available only on rpmfusion:

sudo dnf install https://download1.rpmfusion.org/free/fedora/rpmfusion-free-release-$(rpm -E %fedora).noarch.rpm
sudo dnf install libnvinfer-devel libnvinfer-plugin6

Download tensorflow

I want the version 1.0 of tensorflow:

wget https://github.com/tensorflow/tensorflow/archive/v1.15.2.tar.gz
tar zxf v1.15.2.tar.gz
cd tensorflow-1.15.2/

Create a conda environment where I will install the software

conda create -n tensorflow-gpu-1.15.2 python=3.6
conda activate tensorflow-gpu-1.15.2

Install bazel

To build tensorflow, you need bazel which is a compilation tool. The fedora version is too old and the conda version is too early. I try to install 0.24 or 0.26 bazel versions with conda but it did not work. Conda crashes so I download the linux binary file:

wget https://releases.bazel.build/0.26.1/release/bazel-0.26.1-linux-x86_64
chmod +x ./bazel-0.26.1-linux-x86_64
mv bazel-0.26.1-linux-x86_64 $CONDA_PREFIX/bin/bazel

Install dependencies

I follow instructions on tensorflow website and use pip:

python -m pip install pip six numpy wheel setuptools mock 'future>=0.17.1'
python -m pip install keras_applications --no-deps
python -m pip install keras_preprocessing --no-deps

Build tensorflow from sources and install

Just answer to the questions:

./configure

And launch the long build (maybe hours)

bazel build --config=v1 //tensorflow/tools/pip_package:build_pip_package
./bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg
python -m pip install /tmp/tensorflow_pkg/tensorflow-1.15.2-cp36-cp36m-linux_x86_64.whl

Run the notebook

To run the notebook you need to install some python packages and Keras. Don’t forget to set variables below to be sure that Keras will use the tensorflow backend and the GPU. The value of CUDA_VISIBLE_DEVICES depends on what you get with nvidia-smi. I also create a new jupyter kernel named tf-gpu. Don’t forget to switch to it the first time you execute the notebook.

conda install -y pandas scikit-learn matplotlib
conda install -y gpy gpyopt rdkit
conda install -y keras ipykernel
export CUDA_VISIBLE_DEVICES=0
export KERAS_BACKEND=tensorflow
python -m ipykernel install --user --name tf-gpu
jupyter notebook

Enjoy :-)

links