Joint optimization of DNN model partitioning and slice delivery for distributed edge-cloud inference over optical networks

With the increasing demand for low-latency deep neural network (DNN) inference, edge-cloud collaborative inference has become a promising paradigm. However, the increasing diversity of models, coupled with the limited and heterogeneous resources of edge nodes, makes it impractical to pre-deploy all models at the edge. These constraints not only intensify the complexity of model partitioning and slice delivery but also impose stricter requirements on scheduling and resource coordination. To address these challenges, this paper proposes a joint optimization approach for model partitioning and slice delivery to improve inference performance. We first formulate a mixed-integer nonlinear programming (MINLP) model for exact optimization. A heuristically seeded genetic algorithm (HSGA) is further developed, which incorporates heuristic initialization and task-driven alternating evolution to improve solution quality and convergence speed. Simulation results demonstrate that the proposed algorithm significantly reduces both the task completion time and blocking rate, validating its effectiveness in complex edge-cloud environments.