KV Cache Footprint Evaluation on Context Length 128K 1.0 (test)
2,360.6Effective b_KV (dense)Spherical KV
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Spherical KVModel=Llama-3.1-8B, Context Length (L)=128K2026.05 | 2,360.6 | 1,547.1 | 34.5 | 0.302 | 0.198 | |
| Spherical KVModel=GPT-oss, Context Length (L)=128K2026.05 | 2,442.2 | 1,611.7 | 34 | 0.313 | 0.206 | |
| Spherical KVModel=Qwen2.5-14B, Context Length (L)=128K2026.05 | 2,775.6 | 1,607.7 | 42.1 | 0.355 | 0.206 |