verifiedhigh · 0.90updated 2026-10-04

EAGLE-2 acceptance length and speedup on LLaMA2-Chat 70B (temperature=0)

4.48 tokens_per_cycle

Fields

benchmarks/sd-eagle2-acceptance-llama2-70b-t0

Metric

quality

Value

4.48

Unit

tokens_per_cycle

Model

LLaMA2-Chat 70B

Methodology

temperature=0, MT-bench dataset. Average acceptance length τ = 4.48 tokens per draft-verify cycle. Speedup 3.51x.

Measured by

self

Reproducible

yes

notes

EAGLE-2 with draft tree depth 6, 48 total draft tokens. Draft model is LLaMA2-Chat 7B. Baseline is vanilla autoregressive decoding.

EAGLE-2 with draft tree depth 6, 48 total draft tokens. Draft model is LLaMA2-Chat 7B. Baseline is vanilla autoregressive decoding.