In vehicular ad hoc networks, few existing works on task offloading focus on co-offloading at intra-vehicle level and inter-vehicle level for deep neural network (DNN) inference. Moreover, they ignore the decentralized environment and selfishness of vehicles. This seriously limits the improvement of quality of services for DNN inference. To fill this gap, we formulate a fine-grained offloading problem considering two-level offloading and vehicle selfishness. To solve the problem, we propose a distributed algorithm based on coalition game to encourage vehicles to share their resources. Simulation results show that, the proposed algorithm outperforms the state-of-the-arts in terms of system payoff and running time of algorithms for most cases.