Speculative Decoding on Alpamayo off-shelf (test)
148.2Throughput (tok/s)FlatRoPE
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| FlatRoPEDraft=EAGLE-32026.06 | 148.2 | 7.41 | 4.36 | |
| FlatRoPE + AARLDraft=EAGLE-32026.06 | 145.9 | 7.3 | 4.29 | |
| 3D M-RoPEDraft=EAGLE-32026.06 | 143.8 | 7.35 | 4.23 | |
| FlatRoPEDraft=Dflash2026.06 | 143.1 | 4.95 | 4.21 | |
| FlatRoPE + AARLDraft=Dflash2026.06 | 143.1 | 4.95 | 4.21 | |
| Target onlyDraft=—2026.06 | 34 | — | — |