Generative models have recently been used to produce personalized recommendations by simulating user preferences in continuous space. However, existing work often overlooks crucial discrete structural relationships, such as graph sparsity and user-item topology, reducing recommendation accuracy. Current graph diffusion models that manipulate the full adjacency matrix in continuous space face scalability issues and poorly structured intermediates adjacency matrix. To address these problems, we develop DiffGR, a novel discrete diffusion model for personalised recommendation. DiffGR operates in discrete space, manipulating individual user-item bipartite sub-graphs by randomly removing edges in forward diffusion and restoring them based on a Bernoulli distribution in reverse, preserving graph structure and avoiding full adjacency matrix loading. Empirical studies show that DiffGR achieves competitive personalised recommendation results and generates synthetic graphs with high fidelity to the original dataset, suggesting a promising direction for discrete diffusion modelling in recommender systems.