Test program memory

I found there will too much memory cost in pytorch, somebody give suggestion set cudnn.benchmark = True
to cost down memory, but when I try this in my test program, it warns me
File “/home/.local/lib/python2.7/site-packages/torch/nn/functional.py”, line 52, in conv2d
return f(input, weight, bias)
anyone know how to resolve it?

another problem, in CPU mode, seems the memory cost is higher than GPU mode, is there some way to decrease memory cost?
Thank you!