Multi-physics simulation is essential for addressing complex engineering problems involving interactions between different physical domains. Existing open-source coupling libraries, such as preCICE, have demonstrated good capabilities in partitioned coupling. However, challenges remain in heterogeneous architectures and multi-scale simulations: first, efficient communication between solvers in heterogeneous computing environments is difficult; second, the data format inconsistency between surface meshes and SPH particles can lead to particle penetration when triangle sizes significantly exceed the particle smoothing length, while uniform refinement substantially increases computational costs; third, data mapping between unstructured meshes and meshless particle methods involves complex computations that are difficult to be parallelized efficiently on GPUs. To address these challenges, we propose a heterogeneous communication algorithm and an multi-scale mesh adaptation algorithm for efficient 2-way SPH-FVM and SPH-FEM coupling. Heterogeneous communication algorithms achieve deep overlap between computation and communication by efficiently mapping GPU memory to system memory through a GPU adaptation layer, combined with an asynchronous data exchange mechanism based on multi-streaming and sub-step looping, thereby significantly reducing coupling latency. For SPH-FEM coupling, surface meshes are adaptively refined once during initialization based on the SPH particle search radius, effectively preventing particle penetration while avoiding the efficiency loss caused by excessive refinement. For FVM-SPH coupling, a hierarchical regular grid based on density-adaptive partitioning is employed to establish a one-to-one mapping between airflow particles and grid cells, with volume fractions used to calculate the drag force exerted by the liquid on the airflow, thereby circumventing the complexity of unstructured mesh mapping and its parallelization difficulties. These methods have been integrated into a general-purpose coupling framework and validated through multiple industrial cases using our in-house software PartoX, demonstrating their effectiveness and flexibility.