Time-to-first-token (TTFT) on Llama 8B Instruct 48K input length 3.1
5.2TTFT (s)TriangleMix + FlexPrefill
Evaluation Results
| Method | Links | |
|---|---|---|
| TriangleMix + FlexPrefill2025.07 | 5.2 | |
| FlexPrefill2025.07 | 5.5 | |
| TriangleMix + XAttention2025.07 | 5.5 | |
| DuoAttention2025.07 | 5.8 | |
| TriangleMix2025.07 | 5.9 | |
| TriangleMix + MInference2025.07 | 6 | |
| XAttention2025.07 | 6.6 | |
| Dense2025.07 | 7.3 | |
| MInference2025.07 | 7.8 |