KV Cache Footprint Evaluation on Context Length 8K 1.0 (test)
1,658.3Effective KV Cache Size (dense)Spherical KV
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Spherical KVModel=Llama-3.1-8B, Context Length (L)=8K2026.05 | 1,658.3 | 1,262.3 | 23.9 | 0.013 | 0.01 | |
| Spherical KVModel=Qwen2.5-14B, Context Length (L)=8K2026.05 | 1,690.9 | 1,083.7 | 35.9 | 0.014 | 0.009 | |
| Spherical KVModel=GPT-oss, Context Length (L)=8K2026.05 | 1,717.1 | 1,220.3 | 28.9 | 0.014 | 0.01 |