Long-Context Generation on PG-19 (30K context length)
4.75Throughput Speedup (micro)SpecPV-2K
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SpecPV-2KBackbone=LLaMA3.1-8B-Instruct, Partial KV Budget=2K2025.12 | 4.75 | 3.46 | |
| SpecPV-4KBackbone=LLaMA3.1-8B-Instruct, Partial KV Budget=4K2025.12 | 4.57 | 3.43 | |
| SpecPV-8KBackbone=LLaMA3.1-8B-Instruct, Partial KV Budget=8K2025.12 | 4.23 | 3.42 | |
| EAGLE3-YARNBackbone=LLaMA3.1-8B-Instruct2025.12 | 2.82 | 3.38 | |
| TriForceBackbone=LLaMA3.1-8B-Instruct, Partial KV Budget=4K2025.12 | 2.65 | 2.61 | |
| TokenSwiftBackbone=LLaMA3.1-8B-Instruct, Partial KV Budget=4K2025.12 | 1.83 | 1.73 |