Hi @ptrblck,
Thank you for your comments. I was facing similar issue. Can you please suggest what might be wrong for the following setting?
>>torch.version.cuda
10.2
>>> torch.cuda.is_available()
False
>>> torch.backends.cudnn.enabled
True
$ cat /usr/local/cuda/version.txt
CUDA Version 10.0.130
pytorch version 1.6.0
Can’t play with the main system because Its remote PC and I don’t have rights to upgrade or install NVIDIA driver.
Thank you 
Edit:
I solved this problem by installing cudatoolkit 10.0
I use .yml file with to create conda environment. Now I added following line,
- cudatoolkit=10.0
Conda will select dependency according to this cuda version and hence my pytorch version is also changed, which is ok for me.
Current environment,
>>torch.version.cuda
10.0
>>> torch.cuda.is_available()
True
>>> torch.backends.cudnn.enabled
True
$ cat /usr/local/cuda/version.txt
CUDA Version 10.0.130
pytorch version 1.3.1