Child Module as "weak reference"

1

I have a large pretrained torch.nn.Module that is being used by many different torch.nn.Modules.

For the purpose of example assume it is a pretrained & fixed ResNet image model, that I use for feature generation in many different image classifiers.

How can I best store such model within torch.nn.Module if I do not want the weights to be stored as part od the model state?

Option 1:
Simply storing it as a child module:

class MyModel(torch.nn.Module):
    def __init__(self, resnet: ResNet):
       self.resnet = resnet

Would result in its parameters being stored as part of the Module and increase the checkpoint size.

Option 2

Store the “reference”

class MyModel(torch.nn.Module):
    def __init__(self, resnet: ResNet):
       self._resnet = [resnet]

With this approach the Module just stores the reference to the ResNet and does not really manage it. The syntax is a bit annoying.

Is there a better way to store weak references?
Also, if I want to only enable device management of the child module, but not weight ownership, is overriding apply method the best option?