D3D async diffusion drafter hits 5.71× on Qwen3-32B
TL;DR
- D3D averages 5.71× decoding speedup on Qwen3-32B across seven benchmarks, versus 3.67× for the DFlash diffusion-drafter baseline.
- On GSM8K with Qwen3-32B as target, speedup reaches 7.48×, the highest figure the paper reports.
- Pruning five of 36 drafter layers retains over 99% of mean acceptance length across three benchmarks on Qwen3-8B.
D3D, a standalone diffusion drafter, averages 5.71× decoding speedup with Qwen3-32B as target across seven benchmarks, up from the 3.67× delivered by the DFlash baseline on the same target. Peak on GSM8K: 7.48×.
The authors pin down one specific bottleneck in existing target-conditioned diffusion drafters. The paper reports that "each new block must wait for updated target-model hidden states," which forces drafting and verification to run sequentially. D3D removes that coupling. "This independence lets D3D draft successor blocks during verification in one packed pass," the paper states, which is what makes the decoding asynchronous.
The packed pass is expensive, so the authors prune drafter layers by their effect on acceptance length. With Qwen3-8B as target: "Removing five of 36 layers retains over 99% of mean acceptance length across three benchmarks."
On Qwen3-8B the drafter averages 5.00× versus DFlash's 4.03×. The abstract reports aggregate speedups across seven benchmarks and the GSM8K peak, but does not publish the full per-benchmark breakdown for the 32B target.
Originally reported by paper
Read the original article →Original headline: D3D Hits 5.71× Speculative Decoding on Qwen3-32B—55% Above Prior Best—Via Async Diffusion Drafting