Relational graph convolutional networks (RGCNs) have been successful in learning from knowledge graphs. However, training on large-scale knowledge graphs becomes challenging due to the exponential growth of the neighborhood size across the network layers. Moreover, knowledge graphs have multiple relations, and often, the literals can have multimodal content; these properties make it extra challenging to scale up the training of RGCNs to large-scale graphs. Graph sampling techniques have been shown to be effective in scaling learning to large graphs by reducing the number of processed nodes and lowering memory usage. However, only a few studies have focused on sampling for knowledge graphs. In this work, we introduce ReWise, a relation-wise sampling framework that includes a family of sampling methods designed for knowledge graphs. Our experiments demonstrate that sampling reduces the memory usage up to 50% lower than the case without sampling while maintaining the same classification accuracy and, in some cases, outperforming it. Additionally, we show that our sampling strategy is compatible with the multimodal RGCN, showing the same behavior as RGCNs.