Why MALLET LDA need to keep-sequence?

In the MALLET documentation, it requires --keep-sequence tag for Topic model training (Detail is at : http://ift.tt/11VBte5)

However, in my knowledge, regular LDA modeling use documents as bag of words, since including bigram will increase the feature space by a lot. I wonder why MALLET requires keep-sequence in LDA training, and how did MALLET actually use that sequential information?

Thank you for reading this post.