Research on Transformer Model Compression and Hardware-Friendly Deployment Based on EBSP+GQSA Fusion Method

Existing quantization and sparsity algorithms are basically intended for tradition neural network, which makes them unsuitable to adapt to the Transformer architecture.This paper proposes a dynamically adjustable EBSP+GQSA fusion compression method. We improved the EBSP algorithm by introducing a dynamic sparsity strategy and block-level dynamic shift factors to enhance its adaptability to attention weights. Simultaneously, we optimized the GQSA algorithm by proposing a dynamic grouping strategy and layer-wise weight analysis to improve compression flexibility and efficiency. Experimental results on typical Transformer models, including BERT-base, ViT-base, and GPT-2-small, demonstrate that the EBSP+GQSA fusion method achieves minimal accuracy loss (<1%), the highest compression ratio (>70%), the greatest inference speedup (>2), and the most significant energy reduction (>55%), outperforming the individual use of EBSP or GQSA. This method significantly enhances the deployment capability of models on resource-constrained edge devices, providing a feasible technical path for efficient Transformer inference.