Pt file's size is not constant

Hello everyone,
I am training a U-net model & I save the parameters after each epoch. Recently I noticed that the model size is not constant at all. sometimes it’s a couple of kilobytes & sometimes it’s 112MB!! I save the model after evaluation.
This is how it is saved.

if index == 0:
    model_path = os.path.join(args.models_dir, 'model_{}_topology{}_lr{}_bands{}.pt'.format(model_number, args.model_topology, lr_initial, len(args.bands)))
    torch.save(model.state_dict(), model_path)
    print('log: Saved {}'.format(model_path))
dist.barrier()

Any idea what could be the source of this issue? Or how can approach it?