Long-Context Generation on Qwen3 Context length (60K)
5.89Throughput Speedup (α)SpecPV
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SpecPVModel Size=4B, Partial KV Cache Budget=4K, YARN scaling factor=2.02025.12 | 5.89 | 3.29 | |
| SpecPVModel Size=14B, Partial KV Cache Budget=4K, YARN scaling factor=2.02025.12 | 5.83 | 3.23 | |
| SpecPVModel Size=4B, Partial KV Cache Budget=2K, YARN scaling factor=2.02025.12 | 5.82 | 3.14 | |
| SpecPVModel Size=14B, Partial KV Cache Budget=2K, YARN scaling factor=2.02025.12 | 5.73 | 2.96 | |
| SpecPVModel Size=14B, Partial KV Cache Budget=8K, YARN scaling factor=2.02025.12 | 5.63 | 3.44 | |
| SpecPVModel Size=4B, Partial KV Cache Budget=8K, YARN scaling factor=2.02025.12 | 5.52 | 3.11 | |
| SpecPVModel Size=8B, Partial KV Cache Budget=4K, YARN scaling factor=2.02025.12 | 5.44 | 3.06 | |
| SpecPVModel Size=8B, Partial KV Cache Budget=2K, YARN scaling factor=2.02025.12 | 5.35 | 3.01 | |
| SpecPVModel Size=8B, Partial KV Cache Budget=8K, YARN scaling factor=2.02025.12 | 5.31 | 3.11 | |
| EAGLE3-YARNModel Size=14B, YARN scaling factor=2.02025.12 | 2.98 | 3.31 | |
| EAGLE3-YARNModel Size=8B, YARN scaling factor=2.02025.12 | 2.83 | 3.22 | |
| EAGLE3-YARNModel Size=4B, YARN scaling factor=2.02025.12 | 2.79 | 3.18 |