CODEWITHSUNDEEP
tensorflowgpucublas
Robin
Robin

Reputation: 835

Tensorflow : ImportError: libcublas.so.9.0

I installed Tensorflow. When I try to import it:

import tensorflow as tf

I get the following issue:

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

I have installed CUDA 10 and referenced it in my LD_LIBRARY_PATH:

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda-10.0/lib64

I am working under Ubuntu 18.04.1 with Python 2.7.

Should I uninstall CUDA 10 and install CUDA 9 instead or can I keep working on Tensorflow using CUDA 10 ?

Upvotes: 0

Views: 2042

Answers (3)

Hazarapet Tunanyan
Hazarapet Tunanyan

Reputation: 2865

You don't have to go back for your TensorFlow version. I'm using TensorFlow 1.13 with Cuda 10.1. The problem was related to old installations. I don't have libcublas.so.9.0 in my directories. I'm using Cuda 10.1, but it says the same error. if you've installed the old version previously, make sure you've removed all old versions of TensorFlow and Cuda. you can call pip to uninstall TensorFlow-GPU, it will remove the current, but when you call `pip list | grep TensorFlow, there might be an old version installed on your pc. Remove it completely and install TensorFlow-GPU again. There are 2 issues related to this. take a look at these solutions

ImportError libcublas.so.9.0

Tensorflow 1.13 does not use GPU

Upvotes: 1

Diving
Diving

Reputation: 914

I just had this same problem on Google Colab when trying an GPU accelerator for NuSVR (Thundersvm).

I had to install Cuda 9.0 do overcome it. (Cuda 9.2 , Cuda 10.0 and others, did not worked)

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

credits to Medium Post: Sifat Muhammad Abdullah

Upvotes: 0

GPhilo
GPhilo

Reputation: 19123

Any version of Tensorflow between 1.6 and 1.12 are built against CUDA 9, Tensorflow 1.13 is currently just a release preview, so if you installed via pip install tensorflow-gpu you got the stable version (1.12 as of now).

You then need to install CUDA 9. Note that different CUDA versions can coexist so no need to uninstall CUDA 10

Upvotes: 2

Related Questions