How to add only the parameters that need to be trained to the optimizer？

[Leiyi-Hu]

In my training process, I only need to train a portion of the parameters, and I know that the framework uses requires_grad=False to accomplish parameter freezing, but it seems that the full parameters are still added in the optimizer, so I'm wondering if there is a way to go about adding just the training parameters? Or, to take it a step further, I'd like to know how much of an impact it would have on the memory footprint if I added all of the parameters?