Appears to only be useful for inference. During training, it seems like it is simply just broken.