With the proliferation of artificial intelligence (AI) applications, model partitioning has attracted increasing attention for addressing the computational bottlenecks of edge AI by offloading portions of deep neural networks (DNNs) to cloud servers or IoT devices. However, existing DNN partitioning methods often incur significant parameter overhead during deployment. Moreover, achieving both low-latency inference and privacy preservation in Internet of Things (IoT) networks remains challenging due to device heterogeneity and the heightened privacy risks of distributed architectures. In this work, we propose a reliable distributed computing framework based on fine-grained convolutional kernel partitioning, enabling collaborative execution among multiple edge devices (EDs). To accelerate inference while balancing power and reliability constraints, we jointly optimize kernel allocation, redundancy strategies, and dynamic resource configuration. To address the high coupling of time-scale discrepancies and mixed-variable characteristics, we first apply the Lagrange multiplier method for problem transformation, and then design a two-timescale Hierarchical Proximal Policy Optimization (HPPO) algorithm decoupling the transformed problem into two-layer subproblems. Experimental results show that the proposed framework significantly improves inference speed while ensuring computational reliability in distributed edge environments.