ADAMIX: Adaptive Mixed-Precision Delta-Compression with Quantization Error Optimization for Large Language Models

Abstract: LLMs achieve impressive performance on various knowledge-intensive and complex reasoning tasks in different domains. In certain scenarios like multi-tenant serving, a large number of LLMs finetuned from the same base model are deployed to meet complex requirements for users. Recent works explore delta-compression approaches to quantize and compress the delta parameters between the customized LLM and the corresponding base model. However, existing works either exhibit unsatisfactory performance at high compression ratios or depend on empirical bit allocation schemes. In this work, we propose ADAMIX, an effective adaptive mixed-precision delta-compression framework. We provide a mathematical derivation of quantization error to motivate our mixed-precision compression strategy and formulate the optimal mixed-precision bit allocation scheme as the solution to a 0/1 integer linear programming problem. Our derived bit allocation strategy minimizes the quantization error while adhering to a predefined compression ratio requirement. Experimental results on various models and benchmarks demonstrate that our approach surpasses the best baseline by a considerable margin. On tasks like AIME2024 and GQA, where the norm of $\Delta \mathbf{W}$ is large and the base model lacks sufficient ability, ADAMIX outperforms the best baseline Delta-CoMe by 22.3% and 6.1% with 7B models, respectively.