Speculative Decoding on Specialized Domains Average
114.26Throughput (tokens/s)EvoSpec
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| EvoSpecStatic Vocabulary Size=32k, Dynamic Vocabulary Size=256, Draft Model=Qwen3-1.7B, Target Model=Qwen3-4B, Architecture=Decoupled2026.04 | 114.26 | 1.07 | |
| EvoSpecStatic Vocabulary Size=32k, Dynamic Vocabulary Size=1024, Draft Model=Qwen3-1.7B, Target Model=Qwen3-4B, Architecture=Decoupled2026.04 | 113.11 | 1.06 | |
| EvoSpecStatic Vocabulary Size=16k, Dynamic Vocabulary Size=256, Draft Model=Qwen3-1.7B, Target Model=Qwen3-4B, Architecture=Decoupled2026.04 | 112.63 | 1.05 | |
| EvoSpecStatic Vocabulary Size=32k, Dynamic Vocabulary Size=512, Draft Model=Qwen3-1.7B, Target Model=Qwen3-4B, Architecture=Decoupled2026.04 | 112.58 | 1.05 | |
| EvoSpecStatic Vocabulary Size=16k, Dynamic Vocabulary Size=512, Draft Model=Qwen3-1.7B, Target Model=Qwen3-4B, Architecture=Decoupled2026.04 | 111.15 | 1.04 | |
| EvoSpecStatic Vocabulary Size=16k, Dynamic Vocabulary Size=1024, Draft Model=Qwen3-1.7B, Target Model=Qwen3-4B, Architecture=Decoupled2026.04 | 110.77 | 1.04 | |
| FR-SpecStatic Vocabulary Size=32k, Draft Model=Qwen3-1.7B, Target Model=Qwen3-4B, Architecture=Decoupled2026.04 | 109.08 | 1.02 | |
| Standard SDDraft Model=Qwen3-1.7B, Target Model=Qwen3-4B, Architecture=Decoupled2026.04 | 106.8 | 1 | |
| EvoSpecStatic Vocabulary Size=8k, Dynamic Vocabulary Size=512, Draft Model=Qwen3-1.7B, Target Model=Qwen3-4B, Architecture=Decoupled2026.04 | 104.89 | 0.98 | |
| EvoSpecStatic Vocabulary Size=8k, Dynamic Vocabulary Size=1024, Draft Model=Qwen3-1.7B, Target Model=Qwen3-4B, Architecture=Decoupled2026.04 | 104.12 | 0.97 | |
| EvoSpecStatic Vocabulary Size=8k, Dynamic Vocabulary Size=256, Draft Model=Qwen3-1.7B, Target Model=Qwen3-4B, Architecture=Decoupled2026.04 | 100.9 | 0.94 |