This paper presents a novel dynamic orchestration system for optimizing multi-model composition in autonomous driving scene recognition tasks. Traditional fixed-composition approaches suffer from high latency and suboptimal GPU utilization when handling interdependent models like lane detection, 2D/3D object detection, and vehicle behavior prediction. Our solution introduces a topology-aware vectorization method that abstracts each model’s characteristics into four-dimensional feature vectors. The system dynamically generates optimal execution strategies through a hybrid optimization approach combining genetic algorithms for topological sorting and real-time resource monitoring. Experimental results demonstrate 33% latency reduction compared to serial execution while maintaining full model interoperability. The proposed architecture supports plug-and-play model updates without manual reconfiguration, significantly reducing maintenance overhead in production environments. Benchmarking on NVIDIA 2080ti GPUs shows consistent performance gains across varying workload complexities, with particular advantages in scenarios requiring frequent model updates.