Network predictions are same when the inputs are different, what are the possible reasons?

During training, everything is ok.
However, I got the same prediction results when inputs are differenet. What are the possible reason?
When I use cuda9.2 instead of cuda9.0, the bug dispear.

Are you loading the same model with its state_dict and get constant results for CUDA9.0 and different ones for CUDA9.2?
Could you post a code to reproduce this issue?

I find the reason.

import torch