# Dynamic sequence length in Keras LSTM layer

I'm implementing an LSTM with Keras to predict the correct words order. My dataset is composed by sentences, each sentences is composed by a variable number of words and each word is composed by the same number of features. Now, I would to compute loss function after every sentence parsed, but I need to have dynamic timesteps for each sentence. One solution might be use the fit_generator(), but in the generator I need to know which sentence is to return the correct number of timesteps. Another solution might be use pad_sequences() to a fixed number, but when I want to predict the right label for a word, I have to return to the original sentence length without pad. (for example, If I pad a sentence from 6 words to 20 words, when I predict, the hidden state will be a list of 20 numbers, but I want one number for each original word, so 6 numbers).