'cpu') memory leak issue

I have a workload that has to take a model back and forth to cuda and cpu:

for i in range(iterations):

    with torch.no_grad():'cuda')

    ... operations ...

    with torch.no_grad():'cpu')

This code leads to memory overflow on CPU depending on how many iterations there are.
Am I missing something? or is it a bug?

If I understand the issue correctly your (CPU) RAM is increasing by pushing the model back and forth?
Could you post an executable code snippet, which reproduces this issue so that we could debug it?