EAGLE-1 acceptance length and speedup on Vicuna-13B (temperature=0)
3.82 tokens_per_cycle
Fields
id
benchmarks/sd-eagle1-acceptance-vicuna13b-t0Metric
quality
Value
3.82
Unit
tokens_per_cycle
Model
Vicuna-13B
Methodology
temperature=0, 6 tasks. Average acceptance length τ = 3.82 tokens per draft-verify cycle. Speedup 2.95x mean across tasks.
Measured by
self
Reproducible
yes
notes
EAGLE-1 uses static draft tree with feature-level autoregression. Baseline is vanilla autoregressive decoding. EAGLE-2 improves over EAGLE-1 by 20-40% via context-aware dynamic draft trees.
Notes
EAGLE-1 uses static draft tree with feature-level autoregression. Baseline is vanilla autoregressive decoding. EAGLE-2 improves over EAGLE-1 by 20-40% via context-aware dynamic draft trees.