VirtualEnv tensorflow with Nvidia GPU : cuda-9.0-vs-cuda-8.0, cuDNN-7.0-vs-cuDNN-6.0 - tensorflow

Question

Today I installed tensorflow using the RTFM tensorflow dot org install linux,
I installed the VirtualEnv + Python3 + CPU and tested tensorflow Hello World, which worked fine.
Then I went on the nvidia path (GPU GTX 970) to install VirtualEnv + Python + GPU. RTFM (docs dot nvidia dot com cuda cuda-installation-guide-linux index dot html) , cuda-9.0, cuDDN 7, all PATH's are okay, .bashrc up-to-date, printenv LD_LIBRARY_PATH ok.
My GPU has been ok with the cuda scripts deviceQuery and bandwitdhTest. All post-installation-actions from the Nvidia checklist passed.
When I run Hello World in VirtualEnv+Python3+GPU, code below is what I get (cliffnote : tensorflow wants to use some cudalibrary-8.0 from the /usr/local/cuda-9.0/lib64 which is a 9.0 directory. I tried to add a symlink so cudalibrary-8.0 points to 9.0, but then I get the same message with another cudalibrary... doing this trick for all the cuda libraries is not what I call a fix ;-) )
alexandre#Martin-2:~/Documents/Ordinateur/VirtualEnv$ source tensorflow_py3_gpu/bin/activate
(tensorflow_py3_gpu) alexandre#Martin-2:~/Documents/Ordinateur/VirtualEnv$ python
Python 3.5.2 (default, Sep 14 2017, 22:51:06) [GCC 5.4.0 20160609] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> # Python
... import tensorflow as tf
Traceback (most recent call last):
File "/home/alexandre/Documents/Ordinateur/VirtualEnv/tensorflow_py3_gpu/lib/python3.5/site-packages/tensorflow/python/pywrap_tensorflow.py", line 41, in <module>
from tensorflow.python.pywrap_tensorflow_internal import *
File "/home/alexandre/Documents/Ordinateur/VirtualEnv/tensorflow_py3_gpu/lib/python3.5/site-packages/tensorflow/python/pywrap_tensorflow_internal.py", line 28, in <module>
_pywrap_tensorflow_internal = swig_import_helper()
File "/home/alexandre/Documents/Ordinateur/VirtualEnv/tensorflow_py3_gpu/lib/python3.5/site-packages/tensorflow/python/pywrap_tensorflow_internal.py", line 24, in swig_import_helper
_mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)
File "/home/alexandre/Documents/Ordinateur/VirtualEnv/tensorflow_py3_gpu/lib/python3.5/imp.py", line 242, in load_module
return load_dynamic(name, filename, file)
File "/home/alexandre/Documents/Ordinateur/VirtualEnv/tensorflow_py3_gpu/lib/python3.5/imp.py", line 342, in load_dynamic
return _load(spec)
ImportError: libcublas.so.8.0: cannot open shared object file: No such file or directory
Last line above is about the cudalibrary-8.0 that obviously is NOT in the list of the cudalibrary-9.0. Below is the rest of it.
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "<stdin>", line 2, in <module>
File "/home/alexandre/Documents/Ordinateur/VirtualEnv/tensorflow_py3_gpu/lib/python3.5/site-packages/tensorflow/__init__.py", line 24, in <module>
from tensorflow.python import *
File "/home/alexandre/Documents/Ordinateur/VirtualEnv/tensorflow_py3_gpu/lib/python3.5/site-packages/tensorflow/python/__init__.py", line 49, in <module>
from tensorflow.python import pywrap_tensorflow
File "/home/alexandre/Documents/Ordinateur/VirtualEnv/tensorflow_py3_gpu/lib/python3.5/site-packages/tensorflow/python/pywrap_tensorflow.py", line 52, in <module>
raise ImportError(msg)
ImportError: Traceback (most recent call last):
File "/home/alexandre/Documents/Ordinateur/VirtualEnv/tensorflow_py3_gpu/lib/python3.5/site-packages/tensorflow/python/pywrap_tensorflow.py", line 41, in <module>
from tensorflow.python.pywrap_tensorflow_internal import *
File "/home/alexandre/Documents/Ordinateur/VirtualEnv/tensorflow_py3_gpu/lib/python3.5/site-packages/tensorflow/python/pywrap_tensorflow_internal.py", line 28, in <module>
_pywrap_tensorflow_internal = swig_import_helper()
File "/home/alexandre/Documents/Ordinateur/VirtualEnv/tensorflow_py3_gpu/lib/python3.5/site-packages/tensorflow/python/pywrap_tensorflow_internal.py", line 24, in swig_import_helper
_mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)
File "/home/alexandre/Documents/Ordinateur/VirtualEnv/tensorflow_py3_gpu/lib/python3.5/imp.py", line 242, in load_module
return load_dynamic(name, filename, file)
File "/home/alexandre/Documents/Ordinateur/VirtualEnv/tensorflow_py3_gpu/lib/python3.5/imp.py", line 342, in load_dynamic
return _load(spec)
ImportError: libcublas.so.8.0: cannot open shared object file: No such file or directory
Failed to load the native TensorFlow runtime.
See https tensorflow dot org slash install slash install_sources hashtag common_installation_problems for some common reasons and solutions. Include the entire stack trace above this error message when asking for help.
>>> hello = tf.constant('Hello, TensorFlow!')
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
NameError: name 'tf' is not defined
>>> sess = tf.Session()
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
NameError: name 'tf' is not defined
>>> print(sess.run(hello))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
NameError: name 'sess' is not defined
>>> quit()
(tensorflow_py3_gpu) alexandre#Martin-2:~/Documents/Ordinateur/VirtualEnv$ deactivate`
--
Next day update
Not so clean fix : Create a link in /usr/local/cuda/lib64/ for every library with the wrong number version, linked to the right number version.
alexandre#Martin-2:/usr/local/cuda/lib64$ sudo ln -s libcurand.so.9.0 libcurand.so.8.0
I did this with five cuda libraries (libcusolver, libcublas, libcudart, libcurand, libcufft) and with the cuDNN library libcudnn (version 6 --> version 7).
Hello world! tensorflow worked... but if someone can tell me why tensorflow calls cuda-8.0 and cuDDN-6.0 libraries when I only installed cuda-9.0 and cuDDN-7.0, you're very welcome.
[SOLVED... OR NEARLY] UPDATE
I found https://github.com/tensorflow/tensorflow/issues/12052 that explains almost it all.
Cliffnote :
tensorflow-1.3 uses cuda-8.0 and cuDNN-6.0 (this is why these libraries are linked explicitly when I run tensorflow). I got tricked on the nvidia site that made me download the cuda-9.0 and cuDNN-7.0 versions, which will not be implemented in tensorflow-1.3.
tensorflow-1.4 will work with the cuda-9.0 and cuDNN-7.0 versions. tensorflow-1-4 may be available sometime in October 2017 (or anytime soon, check the link above).

Answer 1

Have you tried
sudo apt install cuda-8-0
?
It should download packages from http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1604
Along with installing cudnn6 (the same way I installed cudnn7), it worked for me.