Fudan's D3D decouples diffusion drafting for asynchronous speculative decoding speedups
Fudan-University · hf · 2026-11-06
Fudan University and collaborators released D3D, a standalone diffusion drafter removing target-model hidden-state dependencies to enable asynchronous speculative decoding. Combining blockwise denoising with auxiliary autoregressive supervision, it drafts successor blocks during verification in one packed pass; pruning 5 of 36 layers retains over 99% of mean accepted length.
More from Infra
- Blue Origin unveils Northstar engine: 715,000 lbf thrust, 30% lighter than BE-4 — beffjezos · 2026-11-06
- DFlash drafts whole token blocks in parallel to fix speculative decoding bottleneck — Lianhuiq · 2026-11-06
- A single desktop B70 runs an entire AI livestream radio, generating music in real time — prince_pringle · 2026-11-06
- Inferact, founded by vLLM creators, now serves 7 trillion tokens a day — Xianbao_QIAN · 2026-11-06
- Q-CTRL launches Dynamics Simulator Qiskit Function on IBM Quantum — MJBiercuk · 2026-11-06
- Data centers projected to cut Mississippi residential electricity rates 16% by 2030 — aronchick · 2026-11-06