Output Equivalence on Qwen3
65.6Exact MatchvLLM + Spec
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| vLLM + SpecSpeculative Decoding=enabled2025.10 | 65.6 | 78.5 | |
| SGLang + EAGLE + DetDrafter=EAGLE, Deterministic=true2025.10 | 50.6 | 69.5 | |
| SGLang + EAGLEDrafter=EAGLE2025.10 | 47.7 | 65.4 | |
| EXSPECBatch Size=42025.10 | 0.95 | 0.971 | |
| Spec-1Batch Size=12025.10 | 0.946 | 0.972 | |
| EQSPECBatch Size=12025.10 | 0.946 | 0.969 | |
| EXSPECBatch Size=12025.10 | 0.946 | 0.969 | |
| Non-Spec-BatchBatch Size=42025.10 | 0.929 | 0.965 | |
| EQSPECBatch Size=42025.10 | 0.923 | 0.957 | |
| BSPBatch Size=12025.10 | 0.035 | 0.199 | |
| BSPBatch Size=42025.10 | 0.021 | 0.126 | |
| DSDBatch Size=12025.10 | 0.002 | 0.022 | |
| DSDBatch Size=42025.10 | 0 | 0.006 |