EAGLE-2 acceptance length and speedup on LLaMA3-Instruct 8B (temperature=0)
4.53 tokens_per_cycle
Fields
id
benchmarks/sd-eagle2-acceptance-llama3-8b-t0Metric
quality
Value
4.53
Unit
tokens_per_cycle
Model
LLaMA3-Instruct 8B
Methodology
temperature=0, MT-bench dataset. Average acceptance length τ = 4.53 tokens per draft-verify cycle. Speedup 3.46x.
Measured by
self
Reproducible
yes
notes
EAGLE-2 on LLaMA3-Instruct 8B. Draft model trained on ShareGPT dataset. Baseline is vanilla autoregressive decoding.
Notes
EAGLE-2 on LLaMA3-Instruct 8B. Draft model trained on ShareGPT dataset. Baseline is vanilla autoregressive decoding.