Production task scheduling is a critical component in achieving large-scale collaborative production in discrete manufacturing. This process not only requires ensuring the efficient transfer of production tasks among geographically dispersed factories but also necessitates the coordinated and flexible cross-plant transfer and allocation of material resources to guarantee timely task completion. Due to the complex inter-factory collaborations, the stochastic nature of task arrivals, and the variable resource demands, especially the additional complexities introduced by cross-plant material transfers, existing scheduling methods struggle to cope effectively, resulting in inefficient coordination and impeding real-time decision-making. To address these challenges, this paper proposes a multi-level attention-based dynamic multi-plant production task scheduling method (MAGMIX) that explicitly considers cross-plant material transfers. First, the dynamic multi-plant production task scheduling problem considering cross-plant material transfers is mathematically formulated; next, the problem is extended to a partially observable Markov decision process; finally, we enhance the classical QMIX algorithm by integrating a Transformer module and a global attention mechanism to effectively capture complex interactions among multiple agents and global contextual information, thereby significantly improving cooperative modeling capabilities within multi-agent systems. Simulation results demonstrate that the proposed MAGMIX algorithm significantly outperforms mainstream multi-agent reinforcement learning algorithms in terms of convergence speed, scheduling rewards, and stability, effectively addressing the challenges posed by continuously dynamic and complex manufacturing environments with cross-plant material transfers.