Fudan's D3D diffusion drafter enables async speculative decoding, up to 7.48x speedup on Qwen3
pmttyji · reddit · 2026-11-07
Fudan University released D3D (Decoupled Diffusion Drafting), a standalone diffusion drafter for asynchronous speculative decoding, with paper, code, and weights all open-sourced.
- Key idea: Target-conditioned diffusion drafters must wait for updated target hidden states between blocks, forcing sequential drafting/verification. D3D combines blockwise denoising with auxiliary autoregressive supervision, removing this dependency so successor blocks are drafted during verification in one packed pass.
- Layer pruning: Layers are pruned by their effect on acceptance length — removing 5 of 36 layers retains over 99% of mean acceptance length.
- Results: Averaging 5.00×/5.71× decoding speedup with Qwen3-8B/Qwen3-32B targets across seven benchmarks (vs. 4.03×/3.67× for DFlash); 7.48× on GSM8K with Qwen3-32B.
- Open resources: Six Qwen3-derived draft checkpoints (1.7B/8B plus pruned variants) and 1.4M-sample training corpora built on Open-PerfectBlend prompts, all on Hugging Face.
More from Infra
- Open-source fork runs 47GiB MoE on 2019 Radeon VII with 49% faster decode — milpster · 2026-11-07
- SmolLM-135M running in pure SQLite at ~40 seconds per token — cneuralnetwork · 2026-11-07
- NUS professor mocks Jensen Huang: he doesn't need a Chinese spy to sell chips to China — pmddomingos · 2026-11-07
- Anthropic's amended S-1 admits mounting losses amid $110B in annual compute commitments — GaryMarcus · 2026-11-07
- Xepak: declarative local builder for REST, RPC, MCP and RAG servers that agents can write DSL for — rumatoest · 2026-11-07
- US AI scale now hinges on coordinating power, components, agentic workflows and humanoid pilots — ingliguori · 2026-11-07