Segmentation fault in torch.svd even for small matrices

I do not know what dependencies are wrong.
I have collected pip wheels from here: torchvision, torch 0.4.0

and installed them. They do not give such segmentation faults.