KV Cache Footprint Analysis on Context Length 32K (test)
1,863.6Effective b_KV (dense)Spherical KV
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Spherical KVModel=Llama-3.1-8B, Context Length (L)=32K2026.05 | 1,863.6 | 1,256.6 | 32.6 | 0.06 | 0.04 | |
| Spherical KVModel=GPT-oss, Context Length (L)=32K2026.05 | 1,927.5 | 1,329.8 | 31 | 0.062 | 0.043 | |
| Spherical KVModel=Qwen2.5-14B, Context Length (L)=32K2026.05 | 2,191.6 | 1,413.4 | 35.5 | 0.07 | 0.045 |