We study the allocation problem in bipartite graphs, in which we wish to minimize the value of some convex objective function subject to the constraints imposed by the graph. Motivated by computational advertising applications, we begin by showing how to implicitly store an optimal matching in space that is much smaller than the size of the graph. We then turn our attention to the online version of the problem. We introduce the online allocation with forecast problem, where we are only given an approximation of the upcoming supply and must construct a compact plan to guide the online allocation. We show that constructing a compact plan on the approximate supply results in a nearly optimal allocation during the run of the online algorithm; these results hold even when the compact plan is constructed using a sampled space.