Compile Tensorflow with GPU Support on Fedora

Since sometimes the training with CPU is low, I try to compile a tensorflow with gpu support. And I install it on a desktop and open openssh-server since my laptop does not have Nvidia GPU.

Environments

Fedora 25 Workstation with GUI.
Kernel version 4.9.14-200.fc25.x86_64.
GTX 950.
We use zsh and oh-my-zsh.
Public IP address.
Need admin permission.

Dependencies

We need compile many things from source code and it is tricky to deal with. Here is the relationship.

gcc5
- compiles by gcc6
Nvidia CUDA and cuDNN
python and pip
- pyenv
- Anaconda
Bazel (compiled from source code)
- gcc
- Java
- others

SSH Support

To install openssh-server, type (Fedora may have it already)

sudo dnf install openssh-server

To see if it is running,

/sbin/service sshd status

To start

systemctl start sshd.service
systemctl enable sshd.service

In order to change the port, just

vim /etc/ssh/sshd_config

and uncomment Port 22 and replace 22 with any port like 12340. Finally,

semanage port -a -t ssh_port_t -p tcp 12340
systemctl restart sshd.service

To test if it works ok,

ssh username@127.0.0.1 -p 12340

Install Gcc 5

By default, Fedora

sudo dnf install gcc gcc-c++ kernel-devel kernel-headers

installs gcc6 and g++. However, CUDA is not compatible with gcc later than 5. Therefore, we need a gcc5. You can get it in another way. Here I choose to compile it from source code. Note that you need gcc5.4 since it seems there is a bug in gcc5.3 when compiling gcc5.3 with gcc6.3.

# Download
wget http://ftp.gnu.org/gnu/gcc/gcc-5.4.0/gcc-5.4.0.tar.gz

# unzip
tar xvfj gcc-5.4.0.tar.gz
cd gcc-5.4.0

# Download prerequisites
./contrib/download_prerequisites
cd ..

# build
mkdir objdir
cd objdir
../gcc-5.4.0/configure --with-system-lib --disable-multilib --enable-languages=c,c++ --prefix=/home/jasonqsy/gcc54
make -j4
make install

Note that you need set prefix by hand since the default prefix is /usr/local. To test it,

~/gcc54/bin/gcc --version

and it shows 5.4.0.

Python version

Personally, I use pyenv to manage versions of python and install anaconda because it is designed for scientific computation. To install pyenv, follow pyenv or type as follows.

git clone https://github.com/pyenv/pyenv.git ~/.pyenv
echo 'export PYENV_ROOT="$HOME/.pyenv"' >> ~/.zshrc
echo 'export PATH="$PYENV_ROOT/bin:$PATH"' >> ~/.zshrc
echo 'eval "$(pyenv init -)"' >> ~/.zshrc
exec $SHELL

And we have pyenv now. To install anaconda, type

pyenv install anaconda3-3.4.0
pyenv global anaconda3-3.4.0

To test the installation, type

conda list
python --version

and should be 3.6. Moreover, protobuf should be installed by

pip install protobuf

Install CUDA

We need

CUDA 8.0
cuDNN

Download CUDA from https://developer.nvidia.com/cuda-downloads. Then

sudo rpm -i cuda-repo-fedora23-8-0-local-ga2-8.0.61-1.x86_64.rpm
sudo dnf clean all
sudo dnf install cuda

and cuda will be installed at /usr/local/cuda-8.0. We will need the directory when compiling tensorflow.

For cuDNN, register and download it from https://developer.nvidia.com/cudnn. Just extracting it is ok. I choose to put it at /usr/local/cudnn.

Install Bazel

Fedora does not have bazel binary support directly. Hence, we need compile from source. Here we use a trick. I download a compiled bazel 0.4.2 to solve the dependencies including java.

Download bazel installer from https://github.com/bazelbuild/bazel/releases. Type

chmod +x bazel-version-installer-os.sh
./bazel-version-installer-os.sh

You can set your custom prefix by —prefix ~. By default, it is /usr/local.