Eagle v2 model for qvikhr 1.5b model. Got following speedup while inferencing using sglang:
| Batch Size | Speedup |
|---|
| 1 | 1.72 |
| 2 | 1.49 |
| 4 | 1.47 |
| 8 | 1.38 |
| 16 | 0.95 |
| 32 | 0.56 |
For training, you can use this commit from the same repo 2f5a5c5bc457034f671a14cf6ff1da4644b4c4f2. (training scripts adaptation to eagle 3 is still in progress, so main branch is broken)
