Install CUDA 10.0, cuDNN 7.3 and build TensorFlow (GPU) from source on Ubuntu 18.04

After CUDA 10.0 and cuDNN 7.3 release I was really eager to try it on my newly built machine. The problem was that pip package TensorFlow 1.11rc didn’t support latest CUDA version and I needed to build it from source. The whole process was rather painful for me and after finally I got it done I decided to go through all steps again and setup it from scratch on empty Ubuntu machine.

My starting point is a machine with i5–8600 CPU 3.10GHz, 16GB DDR4, GTX 1080 with just installed Ubutnu 18.04 LTS and we are going to:
1. Install CUDA 10 + cuDNN 7.3 + NCCL 2.3.5
2. Install Bazel 0.17.2
3. Build and install TensorFlow 1.11rc
Package versions might be newer in you case but assume that overall steps will remain the same.

1. Install CUDA 10.0 + cuDNN 7.3 + NCCL 2.3.5

1.1 Install CUDA 10.0

Great installation guide can be found on NVIDIA site. However below are the steps that were required in my case for empty machine

1.1.1 Install gcc: sudo apt-get install gcc
1.1.2 Download package. For me it was Linux / x86_64 / Ubuntu / 18.04 /deb (local)
1.1.3 Install CUDA by running following commands
sudo dpkg -i cuda-repo-ubuntu1804–10–0-local-10.0.130–410.48_1.0–1_amd64.deb
sudo apt-key add /var/cuda-repo-10–0-local-10.0.130–410.48/7fa2af80.pub
sudo apt-get update
sudo apt-get install cuda

1.1.4 After installation is complete, add PATH variable by adding following line to the bashrc by running
nano ~/.bashrc
adding
export PATH=/usr/local/cuda-10.0/bin${PATH:+:${PATH}}
at the end of file. Save and exit.

1.1.5 Check driver version and CUDA toolkit to ensure that everything went well
cat /proc/driver/nvidia/version
nvcc -V

1.1.6 You might also want to build CUDA samples and run it. It will take a while. For that you need to jump to CUDA sample directory. For me it is /usr/local/cuda-10.0/samples and run
sudo make.
After that go to built sources
/usr/local/cuda-10.0/samples/bin/x86_64/linux/release
and execute ./deviceQuery and ./bandwidthTest. My results are following:

./deviceQuery results

./bandwidthTest results

After test feel free to remove built samples.

1.2 Install cuDNN 7.3

In order to download cuDNN, login/register on developer.NVIDIA.com. Download Release, Dev versions and samples if needed.
Run following commands in he folders with deb files:
sudo dpkg -i libcudnn7_7.3.0.29–1+cuda10.0_amd64.deb
sudo dpkg -i libcudnn7-dev_7.3.0.29–1+cuda10.0_amd64.deb
sudo dpkg -i libcudnn7-doc_7.3.0.29–1+cuda10.0_amd64.deb

The installation is completed. Let’s verify it by following instructions or run:
cp -r /usr/src/cudnn_samples_v7/ $HOME
cd $HOME/cudnn_samples_v7/mnistCUDNN
make clean && make
./mnistCUDNN

If cuDNN was installed properly you will see a message: Test passed
Feel free to remove copied files from HOME/cudnn_samples_v7

1.3 Install NCCL 2.3.5

NCCL is used to handle calculations on multiple GPU simultaneously. If you will use a single GPU, you can skip this step. Otherwise download Network Installer and run
sudo dpkg -i nccl-repo-<version>.deb
sudo apt install libnccl2 libnccl-dev
That is it. Let’s move to Bazel installation.

2 Install Bazel 0.17.2

Bazel is a tool that will be used for TensorFlow building. Its homepage contains multiple ways of installation. For me a preferred one is using ATL repository. First please ensure you have curl
sudo apt install curl
and install Bazel using
sudo apt-get install openjdk-8-jdk
echo “deb [arch=amd64] http://storage.googleapis.com/bazel-apt stable jdk1.8” | sudo tee /etc/apt/sources.list.d/bazel.list
curl https://bazel.build/bazel-release.pub.gpg | sudo apt-key add -
sudo apt-get update
sudo apt-get install bazel

To verify that the bazel installation is completed, bazel version. You might also see few warnings. That is OK.

bazel version result

3. Build TensorFlow

Great step-by-step guide is placed on TensorFlow page. The recap of it is below:

3.1 Preparation.

3.1.1 Install python3-distutils sudo apt-get install python3-distutils
pip sudo apt install python-dev python-pip # or python3-dev python3-pip
I suggest to build TensorFlow using virtual environment. There are multiple ways to do it. My preferred is to use PyCharm as venv manager

3.1.2 Activate virtual environment and run following commands:
pip install -U pip six numpy wheel mock
pip install -U keras_applications==1.0.5 — no-deps
pip install -U keras_preprocessing==1.0.3 — no-deps

3.2 Download and build.

3.2.1 Download source codes. I cloned git repository. If you haven’t git installed, run sudo apt install git
After that clone repository
git clone https://github.com/tensorflow/tensorflow.git
cd tensorflow

3.2.2 Test it with bazel.
bazel test -c opt — //tensorflow/… -//tensorflow/compiler/… -//tensorflow/contrib/lite/…
Well, it was processing about an hour on my machine. As a result I have about 60 failed test results but it doesn’t impact build process.

3.2.3 Configure TensorFlow build by running ./configure.
In my case I’ve used default values except questions regarding:
Hadoop File System support — NO
Apache Kafka Platform support — NO
CUDA support — YES
CUDA version — 10.0
cuDNN version — 7.3
NCCL — 1.3
You might wondering why NCCL version is not 2.3. Well, I’ve tried but NCCL version default path are different from those that are expected during building. It requires to manually copy nccl.h and libnccl.so.2 to required path. As I’m not using multiple GPU, I left 1.3 version which install automatically into expected directories.

3.2.4 Build and install TensorFlow. Run
./bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg
Be ready to entertain yourself for an hour while building process is in progress. As a result pip package will be placed at /tmp/tensorflow_pkg.
We are almost done — run from your venv and enjoy
pip install /tmp/tensorflow_pkg/tensorflow-version-cp27-cp27mu-linux_x86_64.whl

That is it. TensorFlow with GPU support is up and running!