TensorFlow and GPU Installation Archive

Setup GPU Host Machine

http://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#ubuntu-installation

Remove Existing Nouveau Video Driver

• Create a file at /etc/modprobe.d/blacklist-nouveau.conf with the following contents:

blacklist nouveau
options nouveau modeset=0

• Regenerate the kernel initramfs:

sudo update-initramfs -u

Installing CUDA Toolkit

wget http://developer.download.nvidia.com/compute/cuda/7.5/Prod/local_installers/cuda_7.5.18_linux.run

Perform the following steps to install CUDA and verify the installation.

(http://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#runfile)

• Run the installer silently to install with the default selections (implies acceptance of the EULA):

local-gpu$ sh cuda_7.5.18_linux.run --silent

Set LD_LIBRARY_PATH variables:

local-gpu$ export PATH=/usr/local/cuda-7.5/bin:$PATH
local-gpu$ export LD_LIBRARY_PATH=/usr/local/cuda-7.5/lib64:$LD_LIBRARY_PATH

7. Install a writable copy of the samples then build and run the deviceQuery sample

local-gpu$ cuda-install-samples-7.5.sh ~
local-gpu$ cd ~/NVIDIA_CUDA-Samples_7.5/1_Utilities/deviceQuery
local-gpu$ make
local-gpu$ ./deviceQuery

Installing cuDNN

(Download from nvidia directly - requires registration and approval) area

• Something like the following (not exactly... nvidia changes this sometimes)

tar xvzf cudnn-7.5-linux-x64-v2.tgz
sudo cp cudnn-7.5-linux-x64-v2/cudnn.h /usr/local/cuda/include
sudo cp cudnn-7.5-linux-x64-v2/libcudnn* /usr/local/cuda/lib64
sudo chmod a+r /usr/local/cuda/lib64/libcudnn*

Start GPU-enabled Docker

local-gpu$ <pipeline-install-dir>/bin/docker/docker-start-container-gpu.sh

• Uninstall Default CPU TensorFlow and Install GPU TensorFlow

https://www.tensorflow.org/versions/r0.7/get_started/os_setup.html#pip-installation

• TODO: pip uninstall the CPU version of TensorFlow

pip uninstall ...

• TODO: pip install the GPU version of TensorFlow

pip install ...