Building Tensorflow-GPU from source for RTX 2080

GeForce RTX 2080 RIG

Today lets look at how to build a Tensorflow-GPU pip package for the newly available GeForce RTX 2080. The reason behind compiling a custom build will give us the leverage of all the optimizations that are done during the build as well as the major reason being all the available TF-gpu packages till date are built using CUDA-9 and CuDNN 7.3.1. The major issue being CUDA-9 is not aware of the latest GeForce RTX 2080 and its architecture hence when you run a program with TF as backend it shows segmentation fault.

Although I have custom built with many required installations recently I found a github repository which has a bash script which automates all the bull work for this purpose.

STEP 1: Install the required Graphics Driver and CUDA Toolkit 10

This can be done by downloading the run file for Ubuntu 16.04 from here and for Ubuntu 18.04 from here. Now to install the CUDA 10 Toolkit you have to stop the lightdm service first which can be done by issuing the following commands

CTRL + ALT + F1
sudo service lightdm stop
cd /path/to/the/run/file/
sudo sh cuda_10.0.130_410.48_linux.run

Now after the installation of the CUDA Toolkit 10 and the Driver reboot the system

After successful login into the system, run this command in the terminal

nvidia-smi

The output should look something like this

nvidia-smi output

If that is successful run the following commands to add the set the PATHs

sudo echo >> export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda/lib64
sudo echo >> export PATH=/usr/local/cuda/bin${PATH:+:${PATH}}
sudo echo >> export PATH=/usr/local/cuda/lib64${PATH:+:${PATH}}
sudo echo >> export PATH=~/anaconda3/bin:$PATH
sudo echo >> export LD_LIBRARY_PATH=/usr/local/cuda/extras/CUPTI/lib64:$LD_LIBRARY_PATH
sudo echo >> export PATH="$PATH:$HOME/bin"
sudo echo >> export LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu:$LD_LIBRARY_PATH

Open a new terminal and run this following command:

nvcc -V

The output should look something like this

nvcc -V output

Install the CuDNN 7.3.1 deb file by downloading the file from here (you will have to login/create an account)

Install it by issuing the following command

sudo dpkg -i /path/to/the/deb/file/

Install the NCCL deb file by downloading the file from here (you will have to login/create an account)

Install it by issuing the following command

sudo dpkg -i /path/to/the/deb/file/

If you have reached till this point then proceed to step-2

STEP 2: Installing the prerequisites and building the tf-gpu pip file

Download bezel and prerequisites

sudo apt-get install pkg-config zip g++ zlib1g-dev unzip python3=3.5.1 curl pip3
curl https://github.com/bazelbuild/bazel/releases/download/0.18.0/bazel-0.18.0-installer-linux-x86_64.sh
chmod +x bazel-0.18.0-installer-linux-x86_64.sh
./bazel-0.18.0-installer-linux-x86_64.sh
export PATH="$PATH:$HOME/bin"
sudo echo export PATH="$PATH:$HOME/bin" >> .bashrc

Download the TensorFlow source code and configure the bezel build accordingly

git clone https://github.com/tensorflow/tensorflow.git
cd tensorflow
git checkout r1.11
./configure

Bazel Build (This will take quite some time, sit down sip some coffee)

bazel build --config=opt --config=cuda --cxxopt="-D_GLIBCXX_USE_CXX11_ABI=0" //tensorflow/tools/pip_package:build_pip_package

Build the package. The following builds a .whl package in the /tmp/tensorflow_pkg directory copy this to any place you want.

./bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg

STEP 3: Anaconda and installing custom TF-GPU build

Now install Anaconda and install this .whl package using pip within the virtual env of your choice

Pip may show errors with missing packages and incompatible versions correct them manually and you are ready to go  

yes we are done !

P.S : Clap if this guide was useful. the pre-compiled whl package for tensorflow-gpu-cuda10-cuDnn7.3.1 python 3.5.2 is available here