Without self.dim, NetworkModule.calc_scale returns 1.0 even when alpha is present, silently dropping the trained scaling. Use w1b.shape[0] as the rank, matching the network_lora and network_hada conventions.