Since the merge enabling torch.compile, the line self.apply (init.weight_init) in WorldModel.__init__ () does not initialize the Q networks layers since self._Qs.params is not a nn.ParameterList ...