RNNLearner / ULMFIT: Raw outputs and outputs of lower LSTM layers

Hi, what is the exact use of passing along raw_outputs, and outputs for each of the 3 LSTM layers? The both PoolingLinearClassifier and LinearDecoder only seem to need outputs[-1] (and the mask) to proceed. I cannot find where raw_outputs and the two other dimensions for the other LSTM layers of outputs are used.