Accurate optical flow in occluded regions relies heavily on long range motion aggregation, yetthe prevailing attention based GMA module suffers from quadratic complexity and heavy memory footprints, limiting their deployment on resource-constrained edge devices. To address this challenge, we propose SSM-GMA, a novel plug-and-play module that replaces the self-attention mechanism in GMA with four-directional State Space Models (SSMs). By scanning motion feature maps along horizontal, vertical, and diagonal paths, SSM-GMA achieves linear computational complexity while maintaining global receptive fields for effective occlusion handling. The aggregated motion features are fed into a lightweight GRU decoder for iterative flow refinement. Extensive experiments on standard benchmarks demonstrate that SSM-GMA achieves 1.37 px EPE on Sintel Clean and 4.85% F1-all on KITTI-2015, outperforming the original GMA while reducing inference latency (69.2 ms vs. 72 ms). Notably, our method reduces occlusion region error by 4.1% compared to attentionbased approaches, demonstrating superior robustness in challenging scenarios. These results validate that linear-complexity state space dynamics can effectively replace quadratic attention mechanisms for real-time optical flow estimation, enabling practical deployment in embedded systems and edge computing environments.