With the widespread deployment of Graph Neural Networks (GNNs) in edge scenarios, distributed techniques have become common solutions for accelerating GNN inference. Due to platform heterogeneity and differences in computing power, existing solutions generally suffer from problems such as low inference efficiency and workload imbalance. To solve these problems, this paper proposes a GNN inference acceleration framework named HETER-GSD for heterogeneous edge environments. HETER-GSD consists of two modules: (1) HETER, a heterogeneous computing power quantification method based on benchmark testing. By constructing a lightweight test set covering five graphs of diverse topological features, it can effectively evaluate the computing capabilities of different devices on handling GNN inference tasks. (2) GSD, a graph partitioning algorithm adopting a hierarchical seed progressive allocation strategy. GSD reduces the cross-partition communication latency of GNN inference, improves partitioning efficiency, and decreases memory consumption. To verify the effectiveness of the framework, we conduct experiments in a real edge environment composed of five heterogeneous devices, using eight datasets with different scales and characteristics. The results show that, compared with METIS, which is the default partitioning method used by the popular GNN processing libraries PyG and DGL, HETER-GSD achieves a more than 50% improvement in GNN inference time on all datasets while maintaining on-par inference accuracy, demonstrating its superiority in heterogeneous edge GNN inference.