When compiling PyTorch from sources (a80a46a) with the flag NO_DISTRIBUTED=1 it ends up with the Following error :

couldn't find libcudart. Make sure CUDA libraries are installed in adefault location, or that they're in LD_LIBRARY_PATH comming from there : https://github.com/pytorch/pytorch/blob/master/torch/cuda/__init__.py#L66

Compiling with Distributed support makes the error disappear.

Happened on Fedora 29, using GCC/G++ 7.3.1 with CUDA 9.2 / CuDNN 7.4.1 / GTX 1080 Ti