More good news to report! Just heard from our collaborators at Kobo: Curtis Ge and Darius Braziunas that their code contribution is cleared. Kobo released a Latent Dirichlet Allocation (LDA) code which computes inference based on GraphLab LDA model.
In more detail, GraphLab LDA trains the model and classifies existing documents into clusters but when new documents are obtained there is no way for assigning them to the learned classes. Kobo's code classifies new documents based on GraphLab LDA model.
Anyone who wants to test Kobo's code is welcome to try it out. We will soon merge this code into the main GraphLab branch.