Pytorch Nightly, breaks torchvision

I installed pytorch nightly for cuda10 using pip3 by running:

pip3 install torch_nightly -f https://download.pytorch.org/whl/nightly/cu100/torch_nightly.html

Now whenever I try to import torchvision I get:

from torchvision import _C
ImportError: libcudart.so.9.0: cannot open shared object file: No such file or directory

Cuda 10 is installed on my system and is symlinked to /usr/local/cuda/lib64
Here is the output of nvidia-smi

What went wrong? Why is torchvisions looking for cuda 9.0 files?

This might be related to this issue.
I’m not sure, if there is a better way (torchvision-nightly?), but it seems if you would like to use the PyTorch nightly binaries, you should build torchvision from source (or use a stable PyTorch binary).

1 Like

Seems like there is a torchvision nightly as discussed here. I should install that one with pytorch nightly, yes?

The mentioned package was maintained by @stas and, if I’m not mistaken, wasn’t updated anymore, so you would have to wait for the next official torchvision-nightly release or build from source.

1 Like