The stochastic and latent block models are clustering and coclustering tools that are commonly used for network analyses, such as community detection or collaborative filtering. We present a variational inference algorithm for the stochastic block model and the latent block model for sparse graphs, which leverages on the sparsity of edges to scale up to a very large number of nodes. This algorithm is implemented in SparseBM, a Python module that takes advantage of the hardware speed up provided by graphics processing units (GPU).