y_liu
(y_liu)
October 14, 2018, 8:43am
1
Seeking help on “install from source”
I m trying to install from source by building with CUDA10 library. My CUDA and cuDNN installation are working fine and the “python setup.py install” step seems finished correctly (no failure message - correctly identified anaconda python / CUDA paths). But I did not find the torch package installed anywhere (i.e. pytorch folder was not created in python/lib/site-packages ). I wonder if I missed any steps after running setup.py pls?
Thank you!
Just a heads-up.
I am not sure if pytorch is already supported with CUDA10. I tried it once recently and some stuffs did not work.
https://discuss.pytorch.org/t/solved-weird-behavior-with-cuda-toolkit-version-10-0-gcc-6-4-0/26870/3
Why not try with CUDA9 / CUDA9.2?
y_liu
(y_liu)
October 14, 2018, 7:05pm
3
Thank Arul for the reply.
I have got a RTX card so not sure if it works with CUDA9.2
have you solved building pytorch from source with cuda10?
We are building PyTorch from source with CUDA10 pretty often.
Are you running into some issues? If so, could you post the error message etc.?
dsuess
August 2, 2019, 1:51am
6
I’ve used vast.ai’s docker image which includes Pytorch 1.0 with Cuda 10 support without any issues so far.
i’ve tried building pytorch from source with instructions from the link
https://pytorch.org/get-started/locally/
as well as the github repo https://github.com/pytorch/pytorch#from-source
However I didn’t think it was the right instruction since I got the following error when importing torchvision
ImportError: libcudart.so.9.0: cannot open shared object file: No such file or directory
I’m currently running on Ubuntu 16.04 & Python 3.5, without conda.
If you’ve installed torchvision
via pip or conda, it might have accidentally downgraded PyTorch in this process (and expects CUDA9 to be present).
Could you print the current PyTorch version (print(torch.__version__)
) and check, if my assumption is correct?
I would recommend to build torchvision from source after your PyTorch build.
Yes, previously I have installed both packages via pip.
Also I got the following after inserting (print(torch.__version__))
1.2.0a0+2aaeccd
Is the following link still relevant for building torchvision from source?
https://discuss.pytorch.org/t/installing-torchvision-from-source/4677
python setup.py install
should work (instructions ).
It works now. Appreciate the help @ptrblck