A surge of interest in Graph Convolutional Networks (GCN) has produced thousands of GCN variants, with hundreds introduced every year. In contrast, many GCN models re-use only a handful of benchmark datasets as many graphs of interest, such as social or commercial networks, are proprietary. We propose a new graph generation problem to enable generating a diverse set of benchmark graphs for GCNs following the distribution of a source graph -- possibly proprietary -- with three requirements: 1) benchmark effectiveness as a substitute for the source graph for GCN research, 2) scalability to process large-scale real-world graphs, and 3) a privacy guarantee for end-users. With a novel graph encoding scheme, we reframe large-scale graph generation problem into medium-length sequence generation problem and apply the strong generation power of the Transformer architecture to the graph domain. Extensive experiments across a vast body of graph generative models show that our model can successfully generate benchmark graphs with the realistic graph structure, node attributes, and node labels required to benchmark GCNs on node classification tasks.