Stream Generation on StreamChat sequence length (0, 4096]
42.8BLEUFull Attention
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Full Attention2024.07 | 42.8 | 43.8 | 70.9 | |
| ADORE2024.07 | 38.9 | 38.3 | 66.4 | |
| H2O(Rebuilt)2024.07 | 31.2 | 35.7 | 66 | |
| Strided Attention2024.07 | 27.7 | 30.5 | 60.9 | |
| H2O2024.07 | 27.4 | 33.5 | 65.2 | |
| Window Attention2024.07 | 24.7 | 28.5 | 60.6 | |
| KV Compression2024.07 | 21.2 | 32.7 | 61.4 | |
| StreamingLLM2024.07 | 14.6 | 36.1 | 64.8 |