Large-scale computational tasks are increasing in industrial IoT, specifically in control operations. These tasks present computational challenges, and latency becomes a crucial service metric, impacting user experiences and operational efficiency. Addressing these tasks, which exceed the capacity of the local device and cause substantial latency, is essential in industrial IoT. This paper proposes a collaborative offloading mechanism based on terminals and edge servers, aiming to optimize the processing latency in industrial Internet of Things (IIoT) operations. We introduce a hierarchical graph-based collaborative algorithm named Cooperative Graph powered Soft Actor Critic (CoG-SAC), based on Multi-Agent Soft Actor-Critic (MASAC), to address key variables in the optimization model, such as task allocation and resource coordination, for latency minimization. Numerical results show CoG-DDPG has better performance compared with CoG-SAC in improving latency performance.