question about d4 dispersion

[shubbey]

In modules/d4_dispersion_energy.py, we have the following:

        self.register_parameter(
            "_scaleq", nn.Parameter(softplus_inverse(1.0), requires_grad=True)
        )  # for scaling charges of reference systems

Is requires_grad supposed to be True here? I ask because when that is the case, the value of _scaleq updates with each backwards pass. There is a note regarding this in the code,

    def _compute_refc6(self) -> None:
        """
        Function to compute the refc6 tensor. Important: If the charges of
        reference systems are scaled and the scaleq parameter changes (e.g.
        during training), then the refc6 tensor must be recomputed for correct
        results.
        """

But the wording makes it sounds like this would not be the default behavior. For me to get correct results with this term, I do have to manually add the code:

                    loss.backward()
                    model.d4_dispersion_energy._compute_refc6()

in my training loop. Alternatively, if the refc6 term were a parameter with requires_grad=True, I imagine it would properly update itself. Is there a bug here or is this the expected implementation?