Long-context Text Generation on RULER
3.43Mean ThroughputFastKV
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| FastKVBackbone=Qwen2.5-7B, KV Budget=20482026.01 | 3.43 | 2.19 | 0.57 | 0.29 | |
| GemFilterBackbone=Qwen2.5-7B, KV Budget=20482026.01 | 3.32 | 2.2 | 0.41 | 0.14 | |
| FastKVBackbone=Llama-3.1-8B-UL, KV Budget=20482026.01 | 3.07 | 1.84 | 0.49 | 0.36 | |
| GemFilterBackbone=Llama-3.1-8B-UL, KV Budget=20482026.01 | 2.36 | 1.6 | 0.28 | 0.14 | |
| ASLBackbone=Qwen2.5-7B, KV Budget=20482026.01 | 2.35 | 1.54 | 0.4 | 0.21 | |
| ASLBackbone=Llama-3.1-8B-UL, KV Budget=20482026.01 | 2.27 | 0.98 | 0.29 | 0.26 | |
| ASL_2passBackbone=Qwen2.5-7B, KV Budget=20482026.01 | 1.82 | 1.82 | 0.29 | 0.11 | |
| ASL_2passBackbone=Llama-3.1-8B-UL, KV Budget=20482026.01 | 1.57 | 0.7 | 0.18 | 0.09 |