In this paper, we introduce a new post-training compression paradigm for Large Language Models (LLMs) to facilitate their wider adoption. We delve into LLM weight low-rank factorization, and find that the challenges of this task stem from the outlier phenomenon in the LLM activations and the sensitivity difference among various kinds of layers. To address these issues, we propose a training-free approach called Activation-aware Singular Value Decomposition (ASVD). Specifically, ASVD manages activation outliers by scaling the weight matrix based on the activation distribution, thereby enhancing decomposition accuracy. Additionally, we propose an efficient iterative calibration process to optimize layer-specific decomposition by addressing the varying sensitivity of different LLM layers. ASVD can compress a network by 10-20%, without compromising the performance of LLMs. Based on the success of the low-rank decomposition of projection matrices in the self-attention module, we further introduce ASVD to compress the KV cache. By reducing the channel dimension of KV activations, memory requirements for KV cache can be largely reduced. Thanks to the 50-75% reduction in the rank of the KV projection matrices, ASVD can further achieve 50% KV cache reductions without performance drop in a training-free manner.
翻译:暂无翻译