Our project advances federated and split learning for heterogeneous edge–cloud systems, tackling the practical constraints of limited compute, memory, energy, and variable network conditions across client devices. We will design adaptive model partitioning and scheduling that tailor block placement to device profiles, and integrate multi-exit (early-exit) architectures with privacy-preserving, multi-tier secure aggregation to deliver fast, reliable predictions while protecting data. The work entails large-scale multi-GPU training for ablations and hyperparameter sweeps, simulation of diverse client cohorts with realistic churn, and high-throughput storage for checkpoints and federation rounds. Over a one-year period, we will (i) develop profile-guided splitting with online adaptation and (ii) co-train exit heads under federated objectives to minimize end-to-end latency without sacrificing accuracy. Expected outcomes include open-source implementations and reproducible benchmarks spanning vision tasks with non-IID splits.