Abstract
Large language model serving is increasingly limited by KV-cache growth under long-context workloads. Existing KV-cache compression techniques are difficult to compare due to evaluations on different models, tasks, budgets, and serving stacks. This paper presents a workload-aware benchmark of representative KV-cache optimization mechanisms spanning quantization, pruning, and merging, including KIVI, TurboQuant, SnapKV, and CaM.
Benchmark Details
These mechanisms are evaluated on LongBench-style multi-document QA, single-document QA, few-shot learning, and summarization workloads using Llama-3.1-8B-Instruct and Mistral-7B-Instruct-v0.3. The benchmark measures task quality, mean output throughput, mean time-to-first-token, and realized compression ratio across context-length buckets.
Key Findings
Results indicate that the compression ratio alone is a poor predictor of end-to-end performance. KIVI4 provides the most stable quality across models, SnapKV delivers the strongest long-context throughput, and CaM yields significant gains on selected QA workloads but exhibits substantial workload sensitivity in both quality and realized compression ratio. These findings motivate a workload-aware selection of KV-cache mechanisms rather than a one-size-fits-all approach and provide deployment guidance for long-context serving systems.
Blogger's Review: This study systematically showcases the diversity and complexity of KV-cache optimizations, highlighting the importance of selecting strategies tailored to specific workloads. The insights provided will be invaluable for future long-context service developments and should be closely monitored.