Tensorflow: ImportError: libcublas.so.9.0: cannot open shared object file: No such file or directory

I think this is due to the fact that you have CUDA 9.1 and not 9.0, I am facing exactly the same issue.

@Timonzimm I know and I think the whole issue is this f** naming libcublas.so.xxx that nvidia puts. This inherently is mismatch on linux systems whenever that number changes, so since it can not find the exact matches then it thinks the file doesn't exist and throws the error.

I think you should use symbol link from ''cuda/'' to ''cuda/9.1",or your cuda version is too new to tensorflow master branch

@burui11087 I completely forgot about symlinking. Thanks for reminding me.

Seems like this is resolved (and the root cause was a version mismatch of CUDA - 9.0 vs 9.1)?
Closing this out since I understand it to be resolved, but please let me know if I'm mistaken.

FYI @gunan @av8ramit (who are working on the upcoming 1.5 release)

I also occur the exactly same problem with kirk86. For me, I installed cuda toolkit 8.0, and cudnn 5.1.
Then I did what you guys said above, all of them does not work.

For using nightlies, you have to have CUDA 9.0 and cudnn 7 installed.
@yangfengKAUST with the current version of cuda and cudnn installed TF is just complaining that it cannot find the versions it is expecting.

@Timonzimm I am facing the same issue. Have you figured it out?

I have 8.0, 9.0, 9.1 installed + cudnn versions which seem specific to each. The sym linking didn't work from the 9.1 libs. I suspect that sometimes the symlink in the LD_LIBRARY_PATH doesn't work either when I switch versions on the /usr/local/cuda link. I ended up just doing it the low tech way to get the libraries loaded into my java program until I can figure out a cleaner way to handle the paths inside of Eclipse.

    try {
        System.load("/usr/local/cuda/lib64/libcublas.so.9.0");
        System.load("/usr/local/cuda/lib64/libcusolver.so.9.0");
        System.load("/usr/local/cuda/lib64/libcudart.so.9.0");
        System.load("/usr/local/cuda/lib64/libcufft.so.9.0");
        System.load("/usr/local/cuda/lib64/libcurand.so.9.0");

        System.load("/home/greg/Desktop/platform/tensorbuilder/jni/libtensorflow_jni.so");
      } catch (UnsatisfiedLinkError e) {
          System.err.println("Native code library failed to load.\n" + e);
          System.exit(1);
      }

.

@asimshankar Would like to know that in your above comment you mean that we should downgrade cuda to 9.0 and tensorflow 1.5 doesn't work with cuda 9.1 ?

You have closed this issue but its not clear what is the correct action that we should take!

Note: I also have cuda 9.1 installed instead of cuda 9.0.

Just FYI, I have both installed. Building from scratch will work w/ either, but the nightly binaries use 9.0.

@AwasthiMaddy - Yes TensorFlow 1.5 release binaries are built for CUDA 9.

Have you solved it ? This problem is caused tensorflow-gpu-1.5 required cuda 9.0 ,so you should install tensorflow-gpu-1.4. And rember uninstall tensorflow-gpu-1.5. Please use this"pip install --upgrade tensorflow-gpu==1.4"

@aipeteryao - Thank you.

Someone needs to fix the https://www.tensorflow.org/install/install_linux page if this is true, I just followed its instructions exactly, and tells you to install CUDA 8.0 (specifically, not "latest CUDA").

Then as soon as you're done, you get this error (it is looking for cublas 9.0, which, from what I can read here, would not have worked either, as CUDA 9.1 is the default you get from NVIDIA).

Either the webpage instructions should work with the default latest of everything, or it should tell you explicitly to install tensorflow-gpu-1.4 (for example) and not tensorflow-gpu..

Seconding bwesons's comment. I have CUDA 8.0 and Tensorflow 1.3. I followed the current install instructions for TF 1.5 (GPU, ubuntu, virtualenv) and it breaks as described above. Reverting to TF 1.3 until this is resolved.

@aipeteryao This fixed it, thanks! I ended up uninstalling the latest version and installing 1.4, in my virtualenv.

pip3 uninstall tensorflow-gpu
pip3 install --upgrade tensorflow-gpu==1.4

The install page for Ubuntu should be updated: https://www.tensorflow.org/install/install_linux
Since TensorFlow 1.5 is expecting Cuda 9.0 ( NOT 9.1 ), as well as cuDNN 7

In fact, we should view the official document of tensorflow ,it give tensorflow‘s envirment(include python,gcc,cuda,cudnn,an so on).

@bwesen yes,you were right .My computer installed CUDA 8.0,cudnn 6.0 ,tensorflow 1.4.

I think this issue should still be open. @bwesen's comment is correct. The docs tell you to install Cuda 8.0 and use pip install --upgrade tensorflow-gpu. Right now that gives you tensorflow 1.5 which does not work with Cuda 8.0

pinging @asimshankar

I have the same issue (with cuda 9.1 + tensorflow 1.5). I think to resolve it, one option is that to downgrade cuda to 9.0. The other option would be to downgrade both cuda to 8.0 and tensorflow to 1.4. If you have already installed cuda 8.0, you only need to modify LD_LIBRARY_PATH (and CUDA_HOME) environment variable to point to cuda 8.0 directory (i.e. /usr/local/cuda-8.0).

I'm getting this issue (Cuda 9.1.85, cuDNN 7.05)

Tried with tensorflow 1.5, it broke. Uninstalled, installed 1.4 with pip3 install --upgrade tensorflow-gpu==1.4, still broke.

@DylanDmitri 1.5 expects Cuda 9.0, not 9.1

Have you tried with Cuda 9.0 drivers?

@DylanDmitri @mkaze You need Cuda 9.0.

Also, for anyone having trouble installing requirements, I suggest double checking your cuDNN installation. The .deb file didn't work for me because it did not copy files to the right place. I had to use the .tgz file and manually copy files according to nVidia's directions in order to get a working installation.

Why not just install cuda-9-0?

• Go here: https://developer.nvidia.com/cuda-90-download-archive
• Then, for me: Download deb (network)

sudo dpkg -i cuda-repo-ubuntu1604_9.0.176-1_amd64.deb
sudo apt-key adv --fetch-keys \
     http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1604/x86_64/7fa2af80.pub
sudo apt-get update
sudo apt-get install cuda-9-0

Why not just install cuda-9-0?

@abrahamrhoffman That's easy for anyone who has sudo privileges but what about people on a shared system like a cluster environment with simple user privileges. In those cases even if you ask from the sys admin to install any libraries most probably the answer is gonna be NO! Since they are afraid that might interfere with other users' settings and environments.

@abrahamrhoffman Would you also mind providing a justification on the down vote?

I installed cuda-9.0 and still it does not work. This is really irritating.

Please make sure to set your PATH variable appropriately, such as described here: https://stackoverflow.com/questions/39287744/ubuntu-16-04-nvidia-toolkit-8-0-rc-darknet-compilation-error-expected-a/41290056#41290056

export PATH=/usr/local/cuda-9.0/bin${PATH:+:${PATH}}
export LD_LIBRARY_PATH=/usr/local/cuda9.0/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}

cuda 9.1 is the current version. I want TF to use it. How to?

Even tf-nightly-gpu is not looking for cuda 9.1. Meh

I tried symbolic links from all the 9.0 filenames to all the 9.1 filenames and it didn't work. In the end, TF knows the true version. The repo doesn't even have 9.0 anymore so I'm afraid I'll break my nvidia stuff if I remove 9.1 and then manually install 9.0.

I fix him for now by:

Download deb (network) from: https://developer.nvidia.com/cuda-90-download-archive?target_os=Linux&target_arch=x86_64&target_distro=Ubuntu&target_version=1604&target_type=debnetwork

Then: dpkg -i cuda-repo-ubuntu1604_9.0.176-1_amd64.deb

Then: aptitude update

Then: aptitude install cuda-9-0

First I've installed tensorflow 1.5, it broke, and I get the following error:
ImportError: libcublas.so.9.0: cannot open shared object file: No such file or directory
then I uninstalled, installed 1.4 with pip install --upgrade tensorflow-gpu==1.4, it did't work, and I get the following error:
ImportError: libcublas.so.8.0: cannot open shared object file: No such file or directory

@xiezhongzhao What version of Cuda are you using? For tensorflow 1.5 you must have installed the Cuda 9.0 and for tensorflow 1.4 you must use cuda 8.0. If the the tensorflow version and cuda version are compatible, then check the environment variables i.e. CUDA_HOMEand LD_LIBRARY_PATH.

@mkaze I used Cuda9.1

@xiezhongzhao Install Cuda 9.0 and you should be fine. Tensorflow 1.5 does not work with Cuda 9.1.

@mkaze Thank you very much

I am also getting this issue and struggling to resolve it.

$ pip3 install tensorflow-gpu
$ python3
>>> import tensorflow
Traceback (most recent call last):
  File "/home/ubuntu/.local/lib/python3.5/site-packages/tensorflow/python/pywrap_tensorflow.py", line 58, in <module>
    from tensorflow.python.pywrap_tensorflow_internal import *
  File "/home/ubuntu/.local/lib/python3.5/site-packages/tensorflow/python/pywrap_tensorflow_internal.py", line 28, in <module>
    _pywrap_tensorflow_internal = swig_import_helper()
  File "/home/ubuntu/.local/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 "/usr/lib/python3.5/imp.py", line 242, in load_module
    return load_dynamic(name, filename, file)
  File "/usr/lib/python3.5/imp.py", line 342, in load_dynamic
    return _load(spec)
ImportError: libcublas.so.9.0: cannot open shared object file: No such file or directory

I installed following these instructions
https://www.tensorflow.org/install/install_linux#nvidia_requirements_to_run_tensorflow_with_gpu_support

I believe I installed the right versions from nvidia. filenames were
cuda-repo-ubuntu1604_9.0.176-1_amd64.deb
and cudnn-9.0-linux-x64-v7.tgz (version 7.0.5)

I set the path as per those instructions on the tensorflow docs and also tried the instructions that abrahamrhoffman gave above.

When I run ldconfig -v I get some 9.0 libs, but do not see libcublas.so.9.0

/usr/local/cuda-9.0/targets/x86_64-linux/lib:
    libaccinj64.so.9.0 -> libaccinj64.so.9.0.176
    libOpenCL.so.1 -> libOpenCL.so.1.0.0
    libnvToolsExt.so.1 -> libnvToolsExt.so.1.0.0
    libcuinj64.so.9.0 -> libcuinj64.so.9.0.176
    libcudart.so.9.0 -> libcudart.so.9.0.176
/usr/local/cuda-9.1/targets/x86_64-linux/lib:
    libnppif.so.9.1 -> libnppif.so.9.1.85
    libcusparse.so.9.1 -> libcusparse.so.9.1.85
    libcusolver.so.9.1 -> libcusolver.so.9.1.85
    libnpps.so.9.1 -> libnpps.so.9.1.85
    libnppial.so.9.1 -> libnppial.so.9.1.85
    libnvgraph.so.9.1 -> libnvgraph.so.9.1.85
    libcuinj64.so.9.1 -> libcuinj64.so.9.1.85
    libaccinj64.so.9.1 -> libaccinj64.so.9.1.85
    libnppicc.so.9.1 -> libnppicc.so.9.1.85
    libcudart.so.9.1 -> libcudart.so.9.1.85
    libnppc.so.9.1 -> libnppc.so.9.1.85
    libnppicom.so.9.1 -> libnppicom.so.9.1.85
    libOpenCL.so.1 -> libOpenCL.so.1.0.0
    libnppig.so.9.1 -> libnppig.so.9.1.85
    libnvToolsExt.so.1 -> libnvToolsExt.so.1.0.0
    libnppidei.so.9.1 -> libnppidei.so.9.1.85
    libcurand.so.9.1 -> libcurand.so.9.1.85
    libnvblas.so.9.1 -> libnvblas.so.9.1.128
    libnvrtc.so.9.1 -> libnvrtc.so.9.1.85
    libnppitc.so.9.1 -> libnppitc.so.9.1.85
    libnppist.so.9.1 -> libnppist.so.9.1.85
    libcublas.so.9.1 -> libcublas.so.9.1.128
    libnppim.so.9.1 -> libnppim.so.9.1.85
/sbin/ldconfig.real: /usr/local/cuda-9.1/targets/x86_64-linux/lib/libcudnn.so.7 is not a symbolic link

    libcudnn.so.7 -> libcudnn.so.7.0.5
    libcufftw.so.9.1 -> libcufftw.so.9.1.85
    libcufft.so.9.1 -> libcufft.so.9.1.85
    libnppisu.so.9.1 -> libnppisu.so.9.1.85
    libnvrtc-builtins.so.9.1 -> libnvrtc-builtins.so.9.1.85

I did not install 9.1, at least not intentionally. This is on a amazon ec2 instance with stock ubuntu 16.04.
nvidia-smi also returns a gpu, this is a g3.4xlarge instance

any guidance is greatly appreciated.

Per the CUDNN guide at:

http://docs.nvidia.com/deeplearning/sdk/cudnn-install/index.html

You need to copy the unpacked files (from the directory you ran $ tar -xzvf cudnn-9.0-linux-x64-v7.tgz or similar) into /usr/local/cuda subdirectories:

$ sudo cp cuda/include/cudnn.h /usr/local/cuda/include
$ sudo cp cuda/lib64/libcudnn* /usr/local/cuda/lib64
$ sudo chmod a+r /usr/local/cuda/include/cudnn.h
/usr/local/cuda/lib64/libcudnn*

Thank you for the reply @entropy43.

I should have been more specific. I did those two cp and the chmod commands after tar. When I look in the folder cuda folder for where I ran the tar command like ls cuda/lib64 I see
libcudnn.so libcudnn.so.7 libcudnn.so.7.0.5 libcudnn_static.a

I tried this section as well from the nvidia doc
2.3.2. Installing from a Debian File

and the tests pass here in 2.4. Verifying

Any advice to get that lib is appreciated

Could someone please summarize where this currently stands? I am using TF1.3 + cuda 8 with no problem. I'd like to upgrade but it seems like the install process for newer versions is completely broken. Advice?

Similar to what yazabazra is asking above:
TF1.6 Ubuntu 16.04
nvcc -V
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2017 NVIDIA Corporation
Built on Fri_Nov__3_21:07:56_CDT_2017
Cuda compilation tools, release 9.1, V9.1.85
Which requited a Nvidia display driver 390+
Critical to see: https://devtalk.nvidia.com/default/topic/1000340/cuda-setup-and-installation/-quot-nvidia-smi-has-failed-because-it-couldn-t-communicate-with-the-nvidia-driver-quot-ubuntu-16-04/post/5243047/#5243047

whelp to add to it all,, After a major amount of hassle I got the Nvidia updated to the newest release see above, as the TF doc indicated that there were bugs in an earlier release..

Now I'm getting the :
ImportError: libcublas.so.9.0: cannot open shared object file: No such file or directory
Failed to load the native TensorFlow runtime.

Which appears to be a mismatch between 9.0 (TF wants) vs 9.1 Which is most current Nvidia.
It would seem better to run with 9.1 but I'd rather avoid building TF from source and it seems that may not fix it anyhow..

Can this combo be made to work with a binary package?
TF 1.6 Cuda 9.1 ??

Further note and caution to those looking here.. after upgrading my Nvidia stuff my older versions of TF in separate conda env's no longer work as the older TF wants : ImportError: libcublas.so.8.0: cannot open shared object file: No such file or directory
so 9.1 won't cut it.. how about specifying greater than?? vs specific versions? just a suggestion.. In the meantime I'm dead in the water..

And this is why availability of a binary that supports 9.1 would be nice: (from the TF1.6 release notes)

Using XLA:GPU with CUDA 9 and CUDA 9.1 results in garbage results and/or
CUDA_ILLEGAL_ADDRESS failures.

Google discovered in mid-December 2017 that the PTX-to-SASS compiler in CUDA 9
and CUDA 9.1 sometimes does not properly compute the carry bit when
decomposing 64-bit address calculations with large offsets (e.g. load [x + large_constant]) into 32-bit arithmetic in SASS.

As a result, these versions of ptxas miscompile most XLA programs which use
more than 4GB of temp memory. This results in garbage results and/or
CUDA_ERROR_ILLEGAL_ADDRESS failures.

A fix in CUDA 9.1.121 is expected in late February 2018. We do not expect a
fix for CUDA 9.0.x. Until the fix is available, the only workaround is to
downgrade to CUDA 8.0.x
or disable XLA:GPU.

Maybe one of the nightlies does it?

Another solution? can one install multiple revisions of Cuda since TF seems to search for specific Rev's? If so, any advice as to how to?

So Ideally I'd be able to to run TF 1.4(which currently requires Cuda 8.0) in one conda environment and TF 1.6 (which currently requires Cuda 9.0) in another?

So I just added sudo apt-get -y install cuda-toolkit-9.0 and I'm up and running with TF1.6

@dartdog after installing cuda-toolkit-9.0, did you face the issue ImportError: libcudnn.so.7: cannot open shared object file: No such file or directory ?

@dartdog

sudo apt-get install cuda-7-0
vim ~/.bashrc
export PATH=/usr/local/cuda-7.0/bin${PATH:+:${PATH}}
export LD_LIBRARY_PATH=/usr/local/cuda7.0/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
export PATH=/usr/local/cuda-9.0/bin${PATH:+:${PATH}}
export LD_LIBRARY_PATH=/usr/local/cuda9.0/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}

This is definitely supposed to be included in the tensorflow documentation, as said by @bwesen. It should also be included in the errors list. Is it possible for us do this in anyway? I think I am going to open a new issue, requesting to either add support for cuda 9.1, or mention that you need cuda 9.0 specifically in the docs, and also include this in the errors list

Ok, guys. I have now opened a new issue at https://github.com/tensorflow/tensorflow/issues/17629.

I was handling with this issue as well. What worked for me with tensorflow-gpu 1.6:

• I downloaded the toolkit from the archive as 9.0 but it got installed as 9.1 (I do not know why...)
• Still not found libcublas.so.9.0
• Run: sudo apt-get install cuda-libraries-9-0 as suggested at the end of the installation instructions.
• The issue seems to have been solved.

If you want to have tensorflow work with your CUDA version, you need to first uninstall it then compile it from source and specify the CUDA version while running ./configure
Detailed information can be found here: https://www.tensorflow.org/install/install_sources

I am trying this (which builds tensorflow manually)link

Might take a while longer but you can define the minor versions this way.

@mldm4 actually, the command sudo apt-get install cuda probably installed 9.1 for you because you also had that in your system. I had the same problem, and I did sudo apt-get install cuda-9-0 to install a specific version (I had also downloaded from the archive).
I think the commad you did (sudo apt-get install cuda-libraries-9-0) also downloads cuda 9.0.
This issue is not solved, as I mentioned at issue 17629, I know the problem is that TensorFlow expects version 9.0, while I had 9.1. The issue is to mention this more clearly in the docs and include it in the common installation problems in the bottom of the docs, or update tensorflow to accept Cuda 9.1.

so no solution yet?

@thread :

*Please read through the posts carefully! The answer is posted.*

It is your job to read the thread, and discover the solution; not simply scroll to the end.

@abrahamrhoffman that's rude.

I just changed my batchrc from cuda-9.1 to just cuda. Then my tensorflow is able of finding the libcublas.so.9.0

just fyi: nvidia website for downloading cuda-9.0 is actually downloading cuda-9.1. https://developer.nvidia.com/cuda-90-download-archive?target_os=Linux&target_arch=x86_64&target_distro=Ubuntu&target_version=1604&target_type=deblocal

@DanlanChen That is probably because you also have 9.1 installed. To install 9.0, in the steps to download, do sudo apt-get install cuda-9-0 instead of sudo apt-get install cuda.

@cubetastic33 I downgraded my tensorflow version and use cuda8.0, thank you anyway.

@DanlanChen but then, I guess it is preffered to use the latest version. So, if you ever want to upgrade, you now know what to do! :smile:

Im facing the same issue, but I am trying to run tensorflow using nvidia-docker. I have cuda-9-0 installed on the host, but when I try to run my docker container I get

ImportError: libcuda.so.1: cannot open shared object file: No such file or directory

@magick93 and all that turn up here!

LISTEN! Anything you need is downgrade your cuda 9.1 -> cuda 9.0. That's it! Just do it (if you downloaded cuda 9.1 before that you can execute following command in your terminal):

sudo apt-get install cuda-9-0 and remove cuda 9.1 by rm -rf.

Btw, don't forget to change $PATH in your ~/.bashrc (9.1 -> 9.0).

Hi @Oktai15

LISTEN! Anything you need is downgrade your cuda 9.1 -> cuda 9.0. That's it! Just do it

Yes, I have done this - many times.

sudo apt-get install cuda-9-0
[sudo] password for anton: 
Reading package lists... Done
Building dependency tree       
Reading state information... Done
cuda-9-0 is already the newest version (9.0.176-1).
0 upgraded, 0 newly installed, 0 to remove and 7 not upgraded.

$ ldconfig -p | grep cuda
    libnvrtc.so.9.0 (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnvrtc.so.9.0
    libnvrtc.so (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnvrtc.so
    libnvrtc-builtins.so.9.0 (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnvrtc-builtins.so.9.0
    libnvrtc-builtins.so (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnvrtc-builtins.so
    libnvgraph.so.9.0 (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnvgraph.so.9.0
    libnvgraph.so (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnvgraph.so
    libnvblas.so.9.0 (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnvblas.so.9.0
    libnvblas.so (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnvblas.so
    libnvToolsExt.so.1 (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnvToolsExt.so.1
    libnvToolsExt.so (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnvToolsExt.so
    libnpps.so.9.0 (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnpps.so.9.0
    libnpps.so (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnpps.so
    libnppitc.so.9.0 (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnppitc.so.9.0
    libnppitc.so (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnppitc.so
    libnppisu.so.9.0 (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnppisu.so.9.0
    libnppisu.so (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnppisu.so
    libnppist.so.9.0 (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnppist.so.9.0
    libnppist.so (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnppist.so
    libnppim.so.9.0 (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnppim.so.9.0
    libnppim.so (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnppim.so
    libnppig.so.9.0 (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnppig.so.9.0
    libnppig.so (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnppig.so
    libnppif.so.9.0 (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnppif.so.9.0
    libnppif.so (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnppif.so
    libnppidei.so.9.0 (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnppidei.so.9.0
    libnppidei.so (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnppidei.so
    libnppicom.so.9.0 (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnppicom.so.9.0
    libnppicom.so (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnppicom.so
    libnppicc.so.9.0 (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnppicc.so.9.0
    libnppicc.so (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnppicc.so
    libnppial.so.9.0 (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnppial.so.9.0
    libnppial.so (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnppial.so
    libnppc.so.9.0 (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnppc.so.9.0
    libnppc.so (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libnppc.so
    libicudata.so.55 (libc6,x86-64) => /usr/lib/x86_64-linux-gnu/libicudata.so.55
    libcusparse.so.9.0 (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libcusparse.so.9.0
    libcusparse.so (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libcusparse.so
    libcusolver.so.9.0 (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libcusolver.so.9.0
    libcusolver.so (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libcusolver.so
    libcurand.so.9.0 (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libcurand.so.9.0
    libcurand.so (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libcurand.so
    libcuinj64.so.9.0 (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libcuinj64.so.9.0
    libcuinj64.so (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libcuinj64.so
    libcufftw.so.9.0 (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libcufftw.so.9.0
    libcufftw.so (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libcufftw.so
    libcufft.so.9.0 (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libcufft.so.9.0
    libcufft.so (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libcufft.so
    libcudnn.so.7 (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libcudnn.so.7
    libcudart.so.9.0 (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libcudart.so.9.0
    libcudart.so.7.5 (libc6,x86-64) => /usr/lib/x86_64-linux-gnu/libcudart.so.7.5
    libcudart.so (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libcudart.so
    libcudart.so (libc6,x86-64) => /usr/lib/x86_64-linux-gnu/libcudart.so
    libcuda.so.1 (libc6,x86-64) => /usr/lib/x86_64-linux-gnu/libcuda.so.1
    libcuda.so (libc6,x86-64) => /usr/lib/x86_64-linux-gnu/libcuda.so
    libcublas.so.9.0 (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libcublas.so.9.0
    libcublas.so (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libcublas.so
    libaccinj64.so.9.0 (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libaccinj64.so.9.0
    libaccinj64.so (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libaccinj64.so
    libOpenCL.so.1 (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libOpenCL.so.1
    libOpenCL.so (libc6,x86-64) => /usr/local/cuda-9.0/targets/x86_64-linux/lib/libOpenCL.so

Yet I cant even run this simple tensorflow script, as it results I get ImportError: No module named tensorflow.python.client

from tensorflow.python.client import device_lib

def get_available_gpus():
    local_device_protos = device_lib.list_local_devices()
    return [x.name for x in local_device_protos if x.device_type == 'GPU']

get_available_gpus()

I had the same issue everybody is having here, installing 9.0 did help but then I got other issues.
Then I followed this and it finally worked for me.

@Oktai15 doesn't rm -rf delete your system? Please be more clear here, because people might try it without going into the required directory, and end up emptying their home folder.

@magick93 your issue seems to be something else, not the CUDA version.

I had same issue, I think I solved by some changing, this combination works for me

Ubuntu 16, cuda 9.0, cudnn 7.0, python 3,5, tensorflow 1.6

If you install new cuda while you still have previous version please make sure to specify the path like this

export PATH=/usr/local/cuda-9.0/bin${PATH:+:${PATH}}
export LD_LIBRARY_PATH=/usr/local/cuda-9.0/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}

@Abduoit your Ubuntu version doesn't really matter. The thing is, TensorFlow 1.6 expects CUDA to be version 9.0, and cuDNN to be version 7.0.4 (yes, the 0.4 _does_ matter)

Traceback (most recent call last):
File "utils.py", line 15, in
import tensorflow as tf
File "/home/sagar/miniconda2/lib/python2.7/site-packages/tensorflow/__init__.py", line 24, in
from tensorflow.python import *
File "/home/sagar/miniconda2/lib/python2.7/site-packages/tensorflow/python/__init__.py", line 49, in
from tensorflow.python import pywrap_tensorflow
File "/home/sagar/miniconda2/lib/python2.7/site-packages/tensorflow/python/pywrap_tensorflow.py", line 74, in
raise ImportError(msg)
ImportError: Traceback (most recent call last):
File "/home/sagar/miniconda2/lib/python2.7/site-packages/tensorflow/python/pywrap_tensorflow.py", line 58, in
from tensorflow.python.pywrap_tensorflow_internal import *
File "/home/sagar/miniconda2/lib/python2.7/site-packages/tensorflow/python/pywrap_tensorflow_internal.py", line 28, in
_pywrap_tensorflow_internal = swig_import_helper()
File "/home/sagar/miniconda2/lib/python2.7/site-packages/tensorflow/python/pywrap_tensorflow_internal.py", line 24, in swig_import_helper
_mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)
ImportError: libcublas.so.9.0: cannot open shared object file: No such file or directory

Failed to load the native TensorFlow runtime.

See https://www.tensorflow.org/install/install_sources#common_installation_problems
sudo apt-get install cuda-7-0
vim ~/.bashrc
export PATH=/usr/local/cuda-7.0/bin${PATH:+:${PATH}}
export LD_LIBRARY_PATH=/usr/local/cuda7.0/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
export PATH=/usr/local/cuda-9.0/bin${PATH:+:${PATH}}
export LD_LIBRARY_PATH=/usr/local/cuda9.0/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PA
for some common reasons and solutions. Include the entire stack trace
above this error message when asking for help.

As @pascalwhoop mentioned, I followed the instructions in here http://www.python36.com/install-tensorflow141-gpu/ to build Tensorflow from source.

Whenever it said cuda 9.0 I changed to 9.1, and cudnn 7.0 I put 7.1.2.

Worked fine, so far!

@SAGGSOC why are you installing cuda 7.0? You need CUDA 9.0 and CuDNN 7.0.4

Everyone who's been having issues: who's up for turning this into a solid docker image that's shared with the community. Rather do a 6gb image pull once that works than DLing 5 versions of CuDNN before stuff works..

https://github.com/pascalwhoop/tf_openailab_gpu_docker

I started a while back but stopped because of shifting project focuses. But I think it's worth pursuing. Keeps the whole trouble of finding a the right combination of 17 moving parts away from most ppl.

Just to clarify a few things for anyone who might stumble on this post. I have in my system installed cuda-8.0, cuda-9.0, cuda-9.1. You don't have to remove anything to make it work with tensorflow. Instead if you are missing cuda-9.0 from your system, as other have already pointed then you'll need to install it that is a prerequisite for tensorflow to work properly.

If you have cuda-9.0 installed on you system and tensorflow is complaining about libcublas.so.9.0 again as others have said expose that during runtime through your LD_LIBRARY_PATH environment variable in your .bashrc make it point to /usr/local/cuda-9.0/lib64. This should be working even for tensorflow 1.7.

What I have tried and failed to accomplish is build from source. For some reason bazel always exits with an error. If you try to build with cuda-9.0/cuda-9.1 and cudnn7 it complains about gcc7. Using gcc5 compilation seems to be working fine but then at the end I always get an error and the build is unsuccessful.

My question is if anyone has managed to compile from source with cuda-9.1/cuda-9.0 without problems?

This worked for me:

Download CUDA Toolkit 9.0 from NVidia previous releases section.
Then:

sudo dpkg -i cuda-repo-ubuntu1604-9-0-local_9.0.176-1_amd64.deb
sudo apt-key add /var/cuda-repo-9-0-local/7fa2af80.pub 
sudo apt-get update
sudo apt-get install cuda-9.0

Notice 9.0 at the last line above.

export LD_LIBRARY_PATH=/usr/local/cuda-9.0/lib64

My setup
Tensorflow 1.7
cuDNN 7.1.2
Ubuntu 16.04

etemiz,
I've got the same setup and it works for me too.

Thank you for your post :)

I had teh same situation. I had cuda 9.1, and tensorflow would not find libraries for cuda 9.0.

I have installed cuda 9.0 with command: sudo apt-get install cuda-libraries-9-0

That solved my problem.

With cuda 8 and 9.0 installed, setting LD_LIBRARY_PATH in .bashrc and .profile not work. So, I set
LD_LIBRARY_PATH=/usr/local/cuda-9.0/lib64 in pycharm Environment variable field and it works.

the above worked for installing on Ubuntu Server 17. namely,

• installing cuda-9.0 (NOT 9.1)
• cuDNN v7.1.2 (Mar 21, 2018) for CUDA 9.0
• everything else according to the official tf installation instructions

much easier than compiling.

be careful conda users. i hit the same problem and was scratching my head for two days, until finally i discovered that local copy of libcudnn.so was used by conda, under:
/miniconda3/lib/libcudnn.so which pointed to libcudnn.so.7 which pointed to libcudnn.so.7.0.5
i don't remember who and how placed it there but pretty much it overloaded the system default libcudnn.so.7.1.2 !!!

once removed, everything works like a charm:
tensorflow 1.7 or 1.8-nightly, cuda-9.1, cudnn-7.1.2 on ubuntu 16.04

Somehow, I solved this by installing:
cuda 9.1 (from package manager),
cudnn 7.1 for 9.1
and from anaconda:
by using this default command 'conda install -c anaconda tensorflow-gpu'
cudatoolkit 9.0,
tensorflow 1.7,
tensorflow-gpu 1.7

I used Antergos linux, GTX 1060 in my PC. It worked as well in my notebook (Xubuntu 18.04, GT 840m). In my notebook i used :
cuda 9.1 (from nvidia ppa), cudnn 7.1 for 9.1 (from nvidia web), and the rest was the same

Thanks @Suananda! It works like magic.

If you have old version of CUDA, the library link may point to the old library even you install the newer CUDA especially if you install it manually. Try delete your old installation, and then sudo ldconfig to update the dynamic links.

No solution yet!?

For anyone that might stumble on this I have released a community wheel of latest tensorflow 1.8.0-rc1 built with cuda 9.1. You can find it here!

I have find the reason is ldconf, ldconfig is a dynamic link library management command whose purpose is to allow the dynamic link library to be usedby the system.

The default ldconf only search /lib and /usr/lib, as well as the library file under the directory listed in the configuration file /etc/ld. so. conf.

so all of this is caused by the dynamic library of CUDA in the installed CUDA path such as : /path/cuda-9.0/lib64 or /path/cuda-9.0/lib. (for eample my CUDA is installed in /usr/local/cuda-9.0)

1. if you install the CUDA manual, then after install, you should add the path of cuda/lib64 to /etc/ld.so.conf file
   sudo echo "/usr/local/cuda-9.0/lib64/" >> /etc/ld.so.conf
   then
   sudo ldconfig

   of course , you can add the path manual, like:
   vim /etc/ld.so.conf
   then add the path '/usr/local/cuda-9.0/lib64' at the end.
   sudo ldconfig
   after the operation, reopen the ipython or pycharm ,
   import tensorflow as tf
   wow, you will enjoy it!

2. if you install the CUDA by command such as 'dpkg -i cuda-repo-ubuntu1604_9.0.176-1_amd64.deb' or others, it may add the cuda lib path to the /etc/ld.so.conf automatically . but to be on the safe side, check the /etc/ld.so.conf and see if the path add to it .

@NYcleaner
On Ubuntu, there is a /etc/ld.so.conf.d/cuda-9-1.conf file containing :
/usr/local/cuda-9.1/targets/x86_64-linux/lib

Is this enough or do I need to add the directory /usr/local/cuda-9.1/lib64/ to it ?

@sebma
you should add the /usr/local/cuda-9.1/lib64/ path to it， the *.so files are in the lib64

sudo bash -c "echo /usr/local/cuda/lib64/ >/etc/ld.so.conf.d/cuda.conf"
sudo ldconfig

from https://gist.github.com/zhanwenchen/e520767a409325d9961072f666815bb8

@mashu Well the other option is that the community provides pre-built wheels. If you read 2-3 threads above you'll also see that mentioned again.

@Suananda Thanks, it works for me in a tensorflow conda environment.
Suggest the official guys to modify the installation guide "https://www.tensorflow.org/install/install_linux#InstallingAnaconda", step 4 of Anaconda installing, from "pip install --ignore-installed --upgrade tfBinaryURL" to "conda install -c anaconda tensorflow-gpu"

Softlink seems not solve this issue:

➜  cuda ls lib64 
libaccinj64.so                libcufftw_static.a     libnppial.so.9.2      libnppig_static.a    libnvblas.so.9.2
libaccinj64.so.9.2            libcuinj64.so          libnppial.so.9.2.88   libnppim.so          libnvblas.so.9.2.88
libaccinj64.so.9.2.88         libcuinj64.so.9.2      libnppial_static.a    libnppim.so.9.2      libnvgraph.so
libcublas_device.a            libcuinj64.so.9.2.88   libnppicc.so          libnppim.so.9.2.88   libnvgraph.so.9.2
libcublas.so                  libculibos.a           libnppicc.so.9.2      libnppim_static.a    libnvgraph.so.9.2.88
libcublas.so.9.0              libcurand.so           libnppicc.so.9.2.88   libnppist.so         libnvgraph_static.a
libcublas.so.9.2              libcurand.so.9.2       libnppicc_static.a    libnppist.so.9.2     libnvrtc-builtins.so
libcublas.so.9.2.88           libcurand.so.9.2.88    libnppicom.so         libnppist.so.9.2.88  libnvrtc-builtins.so.9.2
libcublas_static.a            libcurand_static.a     libnppicom.so.9.2     libnppist_static.a   libnvrtc-builtins.so.9.2.88
libcudadevrt.a                libcusolver.so         libnppicom.so.9.2.88  libnppisu.so         libnvrtc.so
libcudart.so                  libcusolver.so.9.2     libnppicom_static.a   libnppisu.so.9.2     libnvrtc.so.9.2
libcudart.so.9.2              libcusolver.so.9.2.88  libnppidei.so         libnppisu.so.9.2.88  libnvrtc.so.9.2.88
libcudart.so.9.2.88           libcusolver_static.a   libnppidei.so.9.2     libnppisu_static.a   libnvToolsExt.so
libcudart_static.a            libcusparse.so         libnppidei.so.9.2.88  libnppitc.so         libnvToolsExt.so.1
libcufft.so                   libcusparse.so.9.2     libnppidei_static.a   libnppitc.so.9.2     libnvToolsExt.so.1.0.0
libcufft.so.9.2               libcusparse.so.9.2.88  libnppif.so           libnppitc.so.9.2.88  libOpenCL.so
libcufft.so.9.2.88            libcusparse_static.a   libnppif.so.9.2       libnppitc_static.a   libOpenCL.so.1
libcufft_static.a             libnppc.so             libnppif.so.9.2.88    libnpps.so           libOpenCL.so.1.0
libcufft_static_nocallback.a  libnppc.so.9.2         libnppif_static.a     libnpps.so.9.2       libOpenCL.so.1.0.0
libcufftw.so                  libnppc.so.9.2.88      libnppig.so           libnpps.so.9.2.88    stubs
libcufftw.so.9.2              libnppc_static.a       libnppig.so.9.2       libnpps_static.a
libcufftw.so.9.2.88           libnppial.so           libnppig.so.9.2.88    libnvblas.so

Still got:

Traceback (most recent call last):
  File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/pywrap_tensorflow.py", line 58, in <module>
    from tensorflow.python.pywrap_tensorflow_internal import *
  File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/pywrap_tensorflow_internal.py", line 28, in <module>
    _pywrap_tensorflow_internal = swig_import_helper()
  File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/pywrap_tensorflow_internal.py", line 24, in swig_import_helper
    _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)
  File "/usr/lib/python3.6/imp.py", line 243, in load_module
    return load_dynamic(name, filename, file)
  File "/usr/lib/python3.6/imp.py", line 343, in load_dynamic
    return _load(spec)
ImportError: libcublas.so.9.0: cannot open shared object file: No such file or directory

During handling of the above exception, another exception occurred:

when I run my code on the linux environment directly, everything is OK. But when I run on the local pycharm through the remote interpreter, I encounter the problem: ImportError: libcublas.so.9.0: cannot open shared object file: No such file or directory. "Failed to load the native TensorFlow runtime."

export PATH=${PATH}:/usr/local/cuda-9.0/bin
export CUDA_HOME=${CUDA_HOME}:/usr/local/cuda:/usr/local/cuda-9.0
export LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/cuda-9.0/lib64
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda/extras/CUPTI/lib64

if use pycharm - add it to interpreter

what about f** I just only want to use tensorflow1.8 and cuda9.1?

@dongzhuoyao So what's the problem? look at my comment 6 threads above and you'll find your solution there!

I guess the problem has cropped up again with cuda 9.2 and tensorflow-gpu 1.8. In a virtualenv, I get:

>>> import tensorflow as tf
ImportError: libcublas.so.9.0: cannot open shared object file: No such file or directory

My /usr/local/cuda-9.2/lib64 folder has the following libcublas:

$ ls /usr/local/cuda-9.2/lib64 | grep libcublas
libcublas_device.a
libcublas.so
libcublas.so.9.2
libcublas.so.9.2.113
libcublas.so.9.2.88
libcublas_static.a

Versions:

Ubuntu 16.04
cuda 9.2
cudnn 7.1.4
tensorflow-gpu 1.8.0

We'll need an update to tensorflow-gpu to use cuda 9.2.

Also, if I were to downgrade to cuda 9.0, would I have to first remove cuda 9.2 or just install 9.0 straight away? Would I have conflicting installations?

@mebble Here you go: link. That's tf 1.8 wheel for cuda 9.2.

if I were to downgrade to cuda 9.0, would I have to first remove cuda 9.2

Don't downgrade. Install whatever other version you want they'll get installed at /usr/local/cuda-x.x

Thanks! I forgot to mention that im using python 3.5.2 and pip 10.0.1. I think the wheel is for python 3.6 so the install doesn't work. Do you have one for 3.5 as well?

@kirk86 after installing the whl you gave it throws a similar error for libmpi.so.40
I'm on CentOS and K80 GPU, cuda 9.2 and cudnn v7.1

Suggestion

As far as I know you can have sub-packages xxx,yyy,zzz etc.. and install them as follow

pip install mainpackage[xxx]

This way different co-existing back-ends can be provided. Tensorflow can be build with different options, so at least a couple of cuda-toolkit builds could be provided this way.

The whole point of package is to save time of building, but package build for very specific set of libraries that installs fine, but does not work is counter-productive. It would be better off not to have such package in the first place.

@mebble just make a conda virtual environment for python 3.6. Make sure that you also have installed on your system openmpi.

@pavan-08 Install openmpi on your system also nccl 2.x whatever is the latest from nvidia. I've compiled tf with most of the packages and libraries, so it can be used hdfs, kafta, aws, etc. That's why is asking libmpi.so because it's from openmpi library.

With cuda 9.2 and tensorflow-gpu 1.8 I cannot build tensorflow

declared output 'external/local_config_cuda/cuda/cuda/lib/libcudnn.so.7' is a dangling symbolic link

The symlink exists

Versions:

Ubuntu 17.10
cuda 9.2
cudnn 7.1.4
tensorflow-gpu 1.8.0

This works for me (tensorflow-gpu==1.8.0 and cuda version is 9.0, install in anaconda)

export LD_LIBRARY_PATH=LD_LIBRARY_PATH:/usr/local/cuda-9.0/lib64/

suggestion from: https://stackoverflow.com/questions/48428415/importerror-libcublas-so-9-0-cannot-open-shared-object-file

@Jackiexiao Yea, that's what I was going to say. I'm using a workstation, so for me all I needed to do was use CUDA 9.0 instead of 9.2 (since multiple versions of CUDA were installed)

This worked for me:
Because Tensorflow > 1.4 requires CUDA 9, I uninstalled all CUDA versions :

1. sudo rm -rf /usr/local/cuda*
2. dpkg -l | grep cuda- | awk '{print $2}' | xargs -n1 sudo dpkg --purge

Install excatly CUDA 9.0 and CuDNN 7.0.x and reboot.

Please refer combinations of CUDA, CuDNN and Tensorflow.

This error happens majorly due to incorrect version combinations of Nvidia-driver, CUDA, CuDNN and Tensorflow-gpu

Thank u！

| |
杨康
|
|
邮箱：[email protected]
|

签名由 网易邮箱大师 定制

On 08/18/2018 08:00, Dhruv Srivastava wrote:

Please refer combinations of CUDA and CuDNN.

This error happens majorly due to incorrect version combinations of Nvidia-driver, CUDA, CuDNN and Tensorflow-gpu

—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub, or mute the thread.

@dhruvhacks Good. I think people should install the right version of tensorflow-gpu with the right version cuda.

I am on Ubuntu 18.08, and had Cuda 9.0 installed at:

/usr/local/cuda-9.0

I decided to look for the libcublas.so.9.0:

# build search index
updatedb
# find the "missing" file
locate libcublas.so.9.0

That told me the file was in /usr/local/cuda-9.0/lib64/stubs/libcublas.so.9.0, which helped me realize I had two problems: In ~/.bash_profile I had set: export LD_LIBRARY_PATH=/usr/local/cuda-9.0/lib64/, but my file was in the stubs directory, so I needed to use:

export LD_LIBRARY_PATH=/usr/local/cuda-9.0/lib64/stubs/

Finally, libcublas.so.9.0 was owned by root, so I changed the permissions and the owner of the file:

sudo chown -R MY_USER:MY_GROUP /usr/local/cuda-9.0/
sudo chmod -R 777 /usr/local/cuda-9.0/

And all was well!

With cuda 8 and 9.0 installed, setting LD_LIBRARY_PATH in .bashrc and .profile not work. So, I set
LD_LIBRARY_PATH=/usr/local/cuda-9.0/lib64 in pycharm Environment variable field and it works.

Exactly, Thanks so much for solving the problem, I have spent days to crack this up. really appreciate mate

I encounter this problem as well when I use PyCharm to debug my code at remote server(aws deep learning ami). The cuda version is 9.0 and the cudnn is 7.1.4 and tensorflow is 1.10.0(for gpu). I can import tensorflow normally by terminal connected to my server but fail to import tensorflow in the python console of PyCharm. The problem occurs at this time. Therefore, I cannot debug with PyCharm...
My solution is: input this line in terminal after connecting to your server

sudo ldconfig /usr/local/cuda/lib64

And I referenced thisLink

@NYcleaner Thanks a lot! It helps! and i found just use symbolic link also works.
sudo echo "/usr/local/cuda/lib64/" >> /etc/ld.so.conf
sudo ldconfig

This is still an issue with Cuda 9.1 and Tensorflow 1.11

cuda 10.0 tensorflow 1.11

Even better solution - remove tensorflow, install pytorch
Why tensorflow can brake so easily?
Why can't I just install it and run?
Why is it so easy with pytorch and unpredictable with tf? Wtf is wrong with it?

Why not just install cuda-9-0?

• Go here: https://developer.nvidia.com/cuda-90-download-archive
• Then, for me: Download deb (network)

sudo dpkg -i cuda-repo-ubuntu1604_9.0.176-1_amd64.deb
sudo apt-key adv --fetch-keys \
     http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1604/x86_64/7fa2af80.pub
sudo apt-get update
sudo apt-get install cuda-9-0

Under ubuntu 18.04 cuda 9-0 is not (officially) available. It works, though.

Hi all,
I use PyTorch as much as possible, but for a particular project where i need to export a (Keras) model to tensorflowjs, I'm forced to use tf. The only solution which has worked well for me has been to build from source, after installing CUDA from the Ubuntu multiverse, as described here:

https://medium.com/@asmello/how-to-install-tensorflow-cuda-9-1-into-ubuntu-18-04-b645e769f01d

Bonne chance!

CUDA 9 isn't supported on Ubuntu 18.04 according to official website

Tensorflow doesn't support anything above CUDA 9 according to the official website

Can we get an official CUDA 10 support for tensorflow soon?

Even better solution - remove tensorflow, install pytorch
Why tensorflow can brake so easily?
Why can't I just install it and run?
Why is it so easy with pytorch and unpredictable with tf? Wtf is wrong with it?

Stupid solution

I have installed Cuda 10 and the latest version of Tensor Flow, but I have received ImportError: libcublas.so.9.0: cannot open shared object file: No such file or directory, I have made the downgrade and I still receiving this message...

I have installed Cuda 10 and the latest version of Tensor Flow, but I have received ImportError: libcublas.so.9.0: cannot open shared object file: No such file or directory, I have made the downgrade and I still receiving this message...

Could you tell us the OS version and the exact command you issued to "downgrade"?

I tried Deepin 15.8 and Mint 19.1, I'm going to test Windows 10 later. The command pip install --upgrade tensorflow-gpu==version, I tried the master versions: 1.9.0, 1.10.1, 1.11.0 and 1.12.0. The cuda version and the drivers:

@Sphinxs, in my opinion you have two ways: a) to recompile the python wheel locally so it points to the installed cuda version or b) install cuda-9.0 as I did for my 18.04 installation and it's still working after months.

Even better solution - remove tensorflow, install pytorch
Why tensorflow can brake so easily?
Why can't I just install it and run?
Why is it so easy with pytorch and unpredictable with tf? Wtf is wrong with it?

the man has got a point. TF CAN be compiled to work with more modern CUDA versions... why not just offer that to most people with a tensorflow-gpu-cuda110 package or something temporary until the library can figure out the underlying cuda version dynamically?

Even better solution - remove tensorflow, install pytorch
Why tensorflow can brake so easily?
Why can't I just install it and run?
Why is it so easy with pytorch and unpredictable with tf? Wtf is wrong with it?

I am sick of it!!!!

I am sick of it!!!!

This comment does not add anything to the issue. Also no one forces anyone to use TF, but If you really need to and can't get pip packages working, why not to build it yourself? That's what I ended up doing at some point. I saw there are also community supported builds here
https://github.com/tensorflow/tensorflow just scroll down.
Maybe build for your GPU and contribute?

Even better solution - remove tensorflow, install pytorch
Why tensorflow can brake so easily?
Why can't I just install it and run?
Why is it so easy with pytorch and unpredictable with tf? Wtf is wrong with it?

I am sick of it!!!!

This is an issue page. Please, state your issues so they can be fixed, or, even better, help fixing them.

@raphaunix may I ask where this chart comes from?

I got Tensorflow 1.11.0 working by running the following commands:

conda create -n tf python=2
conda activate tf
pip install tensorflow-gpu==1.11
conda install cudatoolkit==9.0

I have find the reason is ldconf, ldconfig is a dynamic link library management command whose purpose is to allow the dynamic link library to be usedby the system.

The default ldconf only search /lib and /usr/lib, as well as the library file under the directory listed in the configuration file /etc/ld. so. conf.

so all of this is caused by the dynamic library of CUDA in the installed CUDA path such as : /path/cuda-9.0/lib64 or /path/cuda-9.0/lib. (for example my CUDA is installed in /usr/local/cuda-9.0)

1.if you install the CUDA manual, then after install, you should add the path of cuda/lib64 to /etc/ld.so.conf file
sudo echo "/usr/local/cuda-9.0/lib64/" >> /etc/ld.so.conf
then
sudo ldconfig
of course , you can add the path manual, like:
vim /etc/ld.so.conf
then add the path '/usr/local/cuda-9.0' at the end.
then update it
sudo ldconfig
after the operation, reopen the ipython or pycharm ,
import tensorflow as tf
wow, you will enjoy it!

if you install the CUDA by command such as 'dpkg -i cuda-repo-ubuntu1604_9.0.176-1_amd64.deb' or others, it may add the cuda lib path to the /etc/ld.so.conf automatically . but to be on the safe side, check the /etc/ld.so.conf and see if the path add to it .

@jabalazs Instead of conda install cudatoolkit==9.0 I tried pip install cudatoolkit==9.0 but this didn't work, do you know why?

I found that my runtime on colab was not using GPU that's how come I got my error

Even better solution - remove tensorflow, install pytorch
Why tensorflow can brake so easily?
Why can't I just install it and run?
Why is it so easy with pytorch and unpredictable with tf? Wtf is wrong with it?

Good solution @dodler ! but I think you are missing the commands to make this great advice easy to follow through xD.

pip uninstall tensorflow
pip install torch

Happy coding!

I faced this same error trying to use Thundersvm to speedup NuSVR with GPUs on Google Colab.
Installing Cuda 9.0 solved the problem to me.
Just follow:

!wget https://developer.nvidia.com/compute/cuda/9.0/Prod/local_installers/cuda-repo-ubuntu1704-9-0-local_9.0.176-1_amd64-deb

!ls # Check if required cuda 9.0 amd64-deb file is downloaded

!dpkg -i cuda-repo-ubuntu1704-9-0-local_9.0.176-1_amd64-deb

!ls /var/cuda-repo-9-0-local | grep .pub

!apt-key add /var/cuda-repo-9-0-local/7fa2af80.pub

!apt-get update

!sudo apt-get install cuda-9.0