To address the issues of autonomous decision-making difficulties and poor coordination in highly dynamic close-range air combat involving multiple UAVs, a distributed dynamic task allocation method based on multi-agent proximal policy optimization(MAPPO) is proposed. Firstly, considering the executable constraints of tasks and the payload constraints of UAVs, a dynamic task allocation model for multi-UAV is established, targeting enemy UAVs and attack tactics as tasks. Then, a task reallocation network is designed, incorporating a seperated state sliding normalization mechanism, action mask mechanism, and attention mechanism. This network effectively addresses the information distortion issues that occur during the state sliding normalization process in the MAPPO algorithm, ensuring that the task allocation process strictly adheres to task constraints. Additionally, it focuses on the selection of attack tactics based on attack target, enabling distributed dynamic task allocation for multi-UAV. Finally, in a 3v3 air combat scenario, our UAVs equipped with the proposed algorithm engage in air combat against enemy UAVs equipped with an expert system, The combat victory rate of our UAVs reaches 98.5%, validating the effectiveness of the proposed method.