Open cjmcmurtrie opened 7 years ago
I believe this is something to do with pointing at anaconda/lib/libgomp
package instead of system packages, but I have not seen this issue before after many installs of PyTorch on Ubuntu 14.04, CUDA 7.5, in Anaconda environment.
Edit:
I did a fresh PyTorch install after deactivating Anaconda. Like this all tests passed.
I have a feeling Anaconda gcc is too old for latest CUDA 8.0 dependencies, but it's not my area of expertise so not sure. Would love you hear your thoughts before I close the issue.
I had the same problem but with /usr/local/lib/python3.5/dist-packages/torch/lib/libgomp.so.1 . I removed it and for now it's working fine, I think it now uses /usr/lib/x86_64-linux-gnu/libgomp.so.1 installed by the system.
This solution was based on this question http://stackoverflow.com/questions/39017485/openmp-linux-gomp-4-0-not-found .
(Ubuntu 16.04, Python 3.5)
Mine was in my homedir, as I'm using virtualenvs, but same idea as @pablotcarreira:
mv ~/.virtualenvs/pat/lib/python3.5/site-packages/torch/lib/libgomp.so.1{,.backup20170703}
Cool.
Happy to see a conda package created, if someone wants to create one by the way :-)
Thanks, @pablotcarreira , worked for me.
Hi Hugh,
I'm attempting an install on a new GTX1070, Ubuntu 16.10, CUDA 8.0, NVIDIA 367 (it took some research to get this working finally with cutorch and cunn).
Installing PyTorch completes like so:
However the following tests fail, complaining that
version
GOMP_4.0' not found` (related to gcc version? I have 4.9 with I thought was correct):And a
load_lua_class
import complains about the same thing:Any suggestions? Thanks!