Time-to-first-token (TTFT) on Llama 8B Instruct 112K input length 3.1
12.7Time-to-First-Token (s)TriangleMix + MInference
Evaluation Results
| Method | Links | |
|---|---|---|
| TriangleMix + MInference2025.07 | 12.7 | |
| TriangleMix + FlexPrefill2025.07 | 13.6 | |
| FlexPrefill2025.07 | 15 | |
| TriangleMix + XAttention2025.07 | 15.1 | |
| MInference2025.07 | 15.9 | |
| DuoAttention2025.07 | 19.1 | |
| TriangleMix2025.07 | 19.1 | |
| XAttention2025.07 | 19.3 | |
| Dense2025.07 | 27.5 |