SlimSpec low-rank draft LM-head for EAGLE-3: Qwen3-30B-A3B at batch 64
1.21 speedup_ratio: 1.21x end-to-end generation speedup over vanilla autoregressive baseline across MT-Bench, HumanEval, GSM8K at temperature=1, batch size 64. BASELINE: standard autoregressive decoding without speculation at batch 64.
Fields
id
benchmarks/sd-slimspec-eagle3-qwen3-30b-b64Metric
speedup_ratio
Value
1.21
Unit
speedup_ratio: 1.21x end-to-end generation speedup over vanilla autoregressive baseline across MT-Bench, HumanEval, GSM8K at temperature=1, batch size 64. BASELINE: standard autoregressive decoding without speculation at batch 64.
Model
Qwen/Qwen3-30B-A3B
Methodology
Batch size 64, temperature = 1.0. Target model: Qwen3-30B-A3B. Drafter: EAGLE-3 with full-vocabulary baseline LM-head. Average speedup across MT-Bench (1.11x), HumanEval (1.21x), GSM8K (1.32x). Hardware: NVIDIA H100 SXM.
Measured by
third_party
Reproducible
yes
notes
Table 26 in arXiv:2605.10453. Demonstrates batch dilution for EAGLE-3 speculative decoding at batch size 64: average speedup compresses from 1.46x at batch 1 to 1.21x at batch 64, while acceptance length stays virtually identical (tau=3.41 at batch 64 vs 3.46 at batch 1). The compression is entirely due to target verification shifting toward compute-bound execution at batch 64.
Notes
Table 26 in arXiv:2605.10453. Demonstrates batch dilution for EAGLE-3 speculative decoding at batch size 64: average speedup compresses from 1.46x at batch 1 to 1.21x at batch 64, while acceptance length stays virtually identical (tau=3.41 at batch 64 vs 3.46 at batch 1). The compression is entirely due to target verification shifting toward compute-bound execution at batch 64.