The loss should explicitely require a gradient to be evaluated, so I would remove the loss.requires_grad = True line. Also, try to rewrite the first line

def get_pedalboard(self):
    highpass_freq = self.highpass_freq.clamp(20, 500) 

Are you sure the

board = Pedalboard([
            HighpassFilter(cutoff_frequency_hz=float(highpass_freq)),
        ])

works with PyTorch autograd?