KV-Cache Compression Benchmarks — Quantization vs Eviction vs Pruning

Key Points

• The aim is to evaluate the effectiveness of different KV-cache compression techniques.
• Conducted benchmarks on quantization, eviction, and pruning methods.
• Analyzed performance metrics across various cache configurations.
• Compared the compression ratios and speed efficiencies.
• Quantization demonstrated superior compression with minimal performance loss.
• Eviction showed moderate effectiveness but reduced response times.
• Pruning provided the least compression but maintained speed.

Abstract