D3D Method Delivers Up to 7.48x LLM Inference Speedup
First merged summary.
D3D achieves up to 7.48x LLM inference speedup by drafting blocks during verification without target hidden states.
D3D is a standalone diffusion drafter that combines blockwise denoising with autoregressive supervision. It drafts successor blocks during verification in one packed pass, eliminating inter-round waits and sequential dependencies. Layer pruning retains over 99% acceptance length while reducing cost. The method averages 5.00x and 5.71x speedup across seven benchmarks with Qwen3-8B and 32B models, reaching up to 7.48x on GSM8K with Qwen3-32B. Public code is available at GitHub.
- Prunes 5 of 36 layers retaining over 99% acceptance length
- Averages 5.00x and 5.71x speedup on seven benchmarks
- Up to 7.48x speedup on GSM8K with Qwen3-32B
- Combines blockwise denoising and autoregressive supervision
- Eliminates sequential dependency in speculative decoding
- Code at github.com/fudan-generative-vision/D3D
What each outlet adds
beyond the shared facts- Hugging Face daily papersStandalone diffusion drafter removes sequential dependency for 5x LLM inference speedup; up to 7.48x on GSM8K with Qwen3-32B; prunes 5 of 36 layers
- arXiv preprintD3D drafts during verification without target hidden states; averages 5.00x and 5.71x across seven benchmarks
Coverage
oldest first · one row per article · UTC- · 5d ago
Hugging Face daily papersFirst report
D3D enables asynchronous speculative decoding
Standalone diffusion drafter removes sequential dependency for 5x LLM inference speedup.
- · 5d ago
arXiv cs.AI / cs.LG / cs.CL
D3D enables asynchronous speculative decoding
D3D drafts during verification without target hidden states for 5.71x speedup.