I see. It seems the looping is a bug in the model weights but there are bugs in detecting various outputs as identified in the PR I linked.