学术论文翻译成中文

### 直译 The rapid advancement of large language models (LLMs) has revolutionized the field of artificial intelligence, yet deploying them on edge devices remains challenging due to their substantial memory and computation requirements [20]. Existing compression techniques, such as quantization and pruning, often result in significant performance degradation, particularly at extreme compression ratios. To address this, we propose a novel hybrid compression framework that integrates structural pruning with low-rank factorization, targeting the Transformer architecture. Our method preserves the semantic integrity of critical layers while aggressively compressing redundant ones, achieving up to 4.8x reduction in model size without notable accuracy loss on standard benchmarks. Furthermore, we introduce an adaptive fine-tuning strategy that leverages knowledge distillation to recover any minor performance drop, ensuring robustness across heterogeneous hardware platforms, including those with limited computational resources like mobile NPUs. #### ### 意译 ``` 大语言模型（LLMs）的飞速发展彻底改变了人工智能领域，但由于其庞大的内存占用和计算需求，在边缘设备上部署它们仍然是一大挑战 [20]。现有的压缩方法，如量化和剪枝，在极端压缩比下往往会导致性能大幅下降。为了应对这一问题，我们提出了一种新型混合压缩框架，将结构剪枝与低秩分解相结合，专门针对 Transformer 架构设计。该方法在保留关键层语义完整性的同时，对冗余层进行激进压缩，在标准基准测试中实现了最高 4.8 倍的模型体积缩减，且准确率未见明显下降。此外，我们还引入了一种自适应微调策略，利用知识蒸馏技术修复任何细微的性能损失，确保模型在不同硬件平台上（包括计算资源受限的移动 NPU）都能保持鲁棒性。 ```