Yun Chen

Boya Xiong, Shuo Wang, Weifeng Ge et al.

Fine-tuning is a crucial process for adapting large language models (LLMs) to diverse applications. 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 weights between the customized LLM and the corresponding base model. However, they exhibit inadequate performance at high compression ratios due to their empirical nature. In this work, we introduce DeltaMix, an adaptive mixed-precision delta-compression framework designed to minimize quantization error in the singular value decomposition (SVD) space without imposing additional assumptions. DeltaMix provides a theoretical justification for the necessity of mixed-precision compression and presents a practical quantization solution that involves solving a 0/1 linear integer programming problem alongside a reconstruction target correction method. Experimental results across multiple models and benchmarks illustrate that DeltaMix consistently outperforms all baseline methods. Notably, on tasks such as AIME2024 and GQA, DeltaMix exceeds the performance of the best baseline, Delta-CoMe, by 22.3\% and 6.1\% for 7B parameter models, respectively.

Yun Chen, Ali Hajiabadi, Romain Poussier et al.

Side-channel attacks are a security exploit that take advantage of information leakage. They use measurement and analysis of physical parameters to reverse engineer and extract secrets from a system. Power analysis attacks in particular, collect a set of power traces from a computing device and use statistical techniques to correlate this information with the attacked application data and source code. Counter measures like just-in-time compilation, random code injection and instruction descheduling obfuscate the execution of instructions to reduce the security risk. Unfortunately, due to the randomness and excess instructions executed by these solutions, they introduce large overheads in performance, power and area. In this work we propose a scheduling algorithm that dynamically reorders instructions in an out-of-order processor to provide obfuscated execution and mitigate power analysis attacks with little-to-no effect on the performance, power or area of the processor. We exploit the time between operand availability of critical instructions (slack) to create high-performance random schedules without requiring additional instructions or static prescheduling. Further, we perform an extended security analysis using different attacks. We highlight the dangers of using incorrect adversarial assumptions, which can often lead to a false sense of security. In that regard, our advanced security metric demonstrates improvements of 34$\times$, while our basic security evaluation shows results up to 261$\times$. Moreover, our system achieves performance within 96% on average, of the baseline unprotected processor.