SlimSpec low-rank draft LM-head for EAGLE-3: Qwen3-30B-A3B at batch 1
1.46 speedup_ratio: 1.46x end-to-end generation speedup over vanilla autoregressive baseline across MT-Bench, HumanEval, GSM8K at temperature=1, batch size 1. BASELINE: standard autoregressive decoding without speculation.
Fields
id
benchmarks/sd-slimspec-eagle3-qwen3-30b-b1Metric
speedup_ratio
Value
1.46
Unit
speedup_ratio: 1.46x end-to-end generation speedup over vanilla autoregressive baseline across MT-Bench, HumanEval, GSM8K at temperature=1, batch size 1. BASELINE: standard autoregressive decoding without speculation.
Model
Qwen/Qwen3-30B-A3B
Methodology
Batch size 1, temperature = 1.0. Target model: Qwen3-30B-A3B. Drafter: EAGLE-3 with full-vocabulary baseline LM-head. Average speedup across MT-Bench (1.18x), HumanEval (1.58x), GSM8K (1.63x). Hardware: NVIDIA H100 SXM.
Measured by
third_party
Reproducible
yes
notes
Table 24 in arXiv:2605.10453. Measures full-vocabulary EAGLE-3 baseline speedup at batch size 1 (1.46x average, average acceptance length tau=3.46). SlimSpec replaces the draft LM-head with low rank r=d/8, cutting drafter LM-head time by ~50% and improving overall speculative speedup by 8-9% relative to full vocab baseline. Contrast with batch 64 in [[sd-slimspec-eagle3-qwen3-30b-b64]].
Notes
Table 24 in arXiv:2605.10453. Measures full-vocabulary EAGLE-3 baseline speedup at batch size 1 (1.46x average, average acceptance length tau=3.46). SlimSpec replaces the draft LM-head with low rank r=d/8, cutting drafter LM-head time by ~50% and improving overall speculative speedup by 8-9% relative to full vocab baseline. Contrast with batch 64 in [[sd-slimspec-eagle3-qwen3-30b-b64]].