Strange cuDNN mismatch error

My code makes use of 2 networks - net-1 and net-2
When I move net-1 to cuda, there is no problem. However, within the same script when I move net-2 to cuda I get an error :
RuntimeError: cuDNN version mismatch: PyTorch was compiled against 7005 but linked against 5105

Even stranger is that when I try to move net-2 to cuda in a python terminal, I don’t get any error.
It is worth noting that I am using net-1 which was written by someone else, and am combining it with my net-2. The whole code works fine on my local machine, but I am getting this error on a server machine.

Any advice would be greatly appreciated.


Is your server running CUDNN 5?

The only module I have loaded is cudnn 7.0, however it does have the option of cudnn 5 (which I have not loaded). I am not sure how to check which version is being used currently

I found a workaround, which doesnt seem ideal. I disabled cudnn here -

and I don’t get the same error.