KV Cache Compression on 5 sequences Qwen2.5-1.5B-Instruct
0.0358KL DivergenceKVSCULPT
Evaluation Results
| Method | Links | |
|---|---|---|
| KVSCULPTCompression Ratio (r)=0.7, Retain zone (m)=256, Optimizer=L-BFGS, Learning rate=0.52026.03 | 0.0358 | |
| KVSCULPTCompression Ratio (r)=0.5, Retain zone (m)=256, Optimizer=L-BFGS, Learning rate=0.52026.03 | 0.0463 | |
| KVSCULPTCompression Ratio (r)=0.3, Retain zone (m)=256, Optimizer=L-BFGS, Learning rate=0.52026.03 | 0.0575 | |
| Joint OptimizationCompression Ratio (r)=0.7, Retain zone (m)=2562026.03 | 0.115 | |
| Select+FitCompression Ratio (r)=0.7, Retain zone (m)=2562026.03 | 0.125 | |
| Attention ScoreCompression Ratio (r)=0.7, Retain zone (m)=2562026.03 | 0.139 | |
| Joint OptimizationCompression Ratio (r)=0.5, Retain zone (m)=2562026.03 | 0.18 | |
| Select+FitCompression Ratio (r)=0.5, Retain zone (m)=2562026.03 | 0.186 | |
| Attention ScoreCompression Ratio (r)=0.5, Retain zone (m)=2562026.03 | 0.204 | |
| Joint OptimizationCompression Ratio (r)=0.3, Retain zone (m)=2562026.03 | 0.224 | |
| Select+FitCompression Ratio (r)=0.3, Retain zone (m)=2562026.03 | 0.233 | |
| Attention ScoreCompression Ratio (r)=0.3, Retain zone (m)=2562026.03 | 0.254 | |
| RandomCompression Ratio (r)=0.7, Retain zone (m)=2562026.03 | 1.54 | |
| RandomCompression Ratio (r)=0.5, Retain zone (m)=2562026.03 | 1.8 | |
| RandomCompression Ratio (r)=0.3, Retain zone (m)=2562026.03 | 2.5 |