Time-to-first-token (TTFT) on Llama-3.1-8B-Instruct 128K input length
14.5TTFT (s)TriangleMix + MInference
Evaluation Results
| Method | Links | |
|---|---|---|
| TriangleMix + MInference2025.07 | 14.5 | |
| TriangleMix + FlexPrefill2025.07 | 16.1 | |
| FlexPrefill2025.07 | 17.8 | |
| TriangleMix + XAttention2025.07 | 17.8 | |
| MInference2025.07 | 18 | |
| XAttention2025.07 | 23.2 | |
| TriangleMix2025.07 | 23.4 | |
| DuoAttention2025.07 | 23.7 | |
| Dense2025.07 | 34.5 |