Multimodal dynamic fusion, with its multi-sensory capabilities and high accuracy, is becoming increasingly critical in virtual reality, smart homes, autonomous driving, and other scenarios. However, its widespread application is limited by large model sizes, large data volumes, and slow inference speeds. Furthermore, the ever-increasing resolution of multimodal data (e.g., high-definition images and precise sensor readings) substantially escalates computational and transmission demands, thereby intensifying these challenges. Fortunately, the development of edge computing has provided new solutions for the edge deployment of multimodal dynamic fusion. However, parallel inference in heterogeneous edge environments still faces three core challenges: significant differences in computational requirements between modalities, strong heterogeneity in hardware resource requirements for sub-tasks, and dynamic changes in fusion strategies. To address these challenges, this paper proposes the Heterogeneous-device oriented Multimodal dynamic Fusion Parallel inference Framework (HEMPF). This framework employs perception-driven fine-grained data-model co-segmentation, coupled with bidirectional analysis and matching of task hardware requirements and device hardware resources, to generate efficient offloading strategies that can be quickly adjusted according to fusion strategy changes. Experimental results demonstrate that, compared to existing solutions, our approach not only further accelerates multimodal dynamic fusion inference but also exhibits significantly enhanced adaptability to dynamic changes.