Just a documentation of my progress and fails.
Got it *somewhat* working

Ditched debian for this project, because Ubuntu is more commonly mentioned in this field, and I'm impatient to get to business.
Ubuntu 16.04 LTS / NVS 5400m

Tensorflow requirements:

The NVIDIA drivers associated with CUDA Toolkit 9.0.:

Copied from https://tecadmin.net/install-latest-nvidia-drivers-ubuntu/
//When tried nvidia.run installed it failed on debian, this repo works great.
//Cuda 9.1 required nvidia version 390 driver, so no official repo (but read somewhere that Ubuntu 17.10 supports version390 and Cuda 9.1, yet to confirm)

Step 1 – Remove Old Nvidia Driver
The first step is to purge currently installed Nvidia drivers so that it does not conflict with the newer versions.
sudo apt-get purge nvidia*

Step 2 – Install Latest Nvidia Driver
Now enable the graphics-drivers PPA to your system. Currently, it supports Ubuntu 18.04 LTS, 17.10, 17.04, 16.04 LTS, and 14.04 LTS operating systems. Keep in mind that its still under testing phase.
sudo add-apt-repository ppa:graphics-drivers/ppa

Execute the below command to install Nvidia graphics driver on your system.
sudo apt update

sudo apt install nvidia-390

Step 3 – Verify Nvidia Driver
After installation reboot your system, So that your desktop load the new Nvidia driver. Now run below command to check Nvidia module.
lsmod | grep nvidia

CUDA Toolkit 9.0.:

Caution: Tried installing .deb version from Cuda 9.1 downloads
Followed the .deb installation recipe:
sudo dpkg -i cuda-repo-ubuntu1604-9-1-local_9.1.85-1_amd64.deb
sudo apt-key add /var/cuda-repo-<version>/7fa2af80.pub
sudo apt-get update
sudo apt-get install cuda

At last step, apt wanted to replace nvidia v390 to v387 when listed dependencies but the cuda page cleary states
Before installing the CUDA Toolkit on Linux, please ensure that you have the latest NVIDIA driver R390 installed.

SO I USED .run installer and it ran without errors

Accidently installed 9.1, which might be the problem why my test sample wont execute... other possible cause "Compute Capability 3.0 required", NVS 5400 is 2.1

NVIDIA graphics driver R375 or newer for CUDA 8
NVIDIA graphics driver R384 or newer for CUDA 9
NVIDIA graphics driver R390 or newer for CUDA 9.1

cuDNN v7.0.

A GPU of compute capability 3.0 or higher is required..
That cannot be satisfied however Tensorflow might run without it. As ive understood cuDNN is needed for various optimizations especially for memory utilization. The card will perform worse without it.

Libcupti-dev library

Copied from: https://www.tensorflow.org/install/install_linux
The libcupti-dev library, which is the NVIDIA CUDA Profile Tools Interface. This library provides advanced profiling support. To install this library, issue the following command for CUDA Toolkit >= 8.0:
sudo apt-get install cuda-command-line-tools

//package didnt exist in Ubuntu 16.04
and add its path to your LD_LIBRARY_PATH environment variable:
export LD_LIBRARY_PATH=${LD_LIBRARY_PATH:+${LD_LIBRARY_PATH}:}/usr/local/cuda/extras/CUPTI/lib64

//wrong export path it was : /usr/local/cuda9.1smth/extras/...

Contributors to this page: null .
Page last modified on Sunday 15 of April, 2018 01:25:22 CEST by null.

Upcoming Events

1)  Tue 01 of Jan, 2019 18:00 CET
Kotes workshop