The discovery of peptide-based drugs is a major focus in drug R&D, with Graph Neural Networks (GNNs) achieving notable success in predicting antimicrobial peptides (AMPs). However, AMP-GNN models face acceleration challenges due to heavy preprocessing and computational overhead. Simplifying preprocessing often compromises performance. To address this, we propose REAMP, a redundancy elimination system tailored for AMP-GNNs. REAMP features a deep redundancy elimination algorithm, matrix-based operator reconstruction, and a heterogeneous GPU-CPU processing pipeline to enhance redundancy removal and reduce overhead on dense graph batches. Extensive experiments show REAMP significantly reduces preprocessing and end-to-end costs. Integrated into TP-LMMSG, it achieves up to 2.3× speedup (average 2.0 ×) and a 4.2 × improvement in redundancy elimination over state-of-the-art methods. Our source code is available at: https://github.com/Anonymous-615/REAMP .