I all, I have another problem when I am trying to use tvm to compile tensorflow framework. The problem is that the system can not find libcublas.so.10.0 when I import tensorflow in python.
First, my computer is CentOS7.6 and I have a GPU card which is nvidia quadro K600. I download cuda toolkit 10.0.1 and the tensorflow I install is tensorflow-gpu 1.13.0. I have checked that tensorflow 1.13.0 support cuda toolkit 10.0, and there is indeed a libcublas.so.10.0 in my /usr/lib64 folder. Has anyone came across the same problem with mine ? Really appreciated for some suggestions.
This doesn’t look issue with TF frontend.
1: Is your tensorflow setup work fine ?
2: Are you able to compile TVM in your environment with CUDA enabled ?
Please share error logs is possible .
Thanks for your reply, I think I find the problem. But the way I solve this problem is by using anaconda virtual environment. I still do not know how to solve this problem if I do not use anaconda.
When I open python, the terminal will shows the python version and the related gcc version. As I previously use pip to install python3, so I think they use the default gcc to build python3, which is 4.8.5.
I install anaconda and create an virtual environment for python3.6, and the default gcc version is changed to 7.3.0. So looks like anaconda automatically bind a new gcc compile tool when it install python3 in its environment.
But I just wonder how it build the gcc 7.3.0. Because I install gcc 8.3.2 instead of gcc 7.3.0 and I can not find gcc lib or bin inside my anaconda virtual environment.
I don’t think it’s a TVM issue and you might be interested in flags like LD_LIBRARY_PATH