D3D
A diffusion drafting model proposed by Fudan University that employs block-level denoising with autoregressive auxiliary supervision to achieve asynchronous speculative decoding, improving the inference speed of large models.
Timeline (last 90 days)
D3D is a speculative decoding method proposed by Fudan University's Generative Vision team that accelerates large model inference by decoupling the drafter's dependence on the target model's hidden states.
D3D is a standalone diffusion drafter whose core design combines block-level denoising (generating an entire block of candidate tokens in parallel) with auxiliary autoregressive supervision (maintaining a longer high-quality accepted prefix).
D3D abandons the use of the target model's hidden states as a conditioning signal and instead compensates drafting quality with autoregressive supervision.
D3D can draft subsequent blocks via a single packed pass while the target model performs verification, enabling asynchronous speculative decoding.
D3D employs layer-wise progressive pruning to reduce the extra computational overhead of asynchronous packed drafting, with each layer pruned according to its impact on acceptance length.
On three benchmarks with Qwen3-8B as the target model, D3D still retains over 99% of the average acceptance length after pruning 5 of 36 layers.
In evaluations across seven benchmarks, D3D achieves an average decoding speedup of 5.00x with Qwen3-8B as the target model.
On the GSM8K benchmark, D3D reaches a speedup of 7.48x with Qwen3-32B as the target model.
D3D's code is open-sourced on GitHub at the repository path fudan-generative-vision/D3D.
All Facts (9)
D3D is a speculative decoding method proposed by Fudan University's Generative Vision team that accelerates large model inference by decoupling the drafter's dependence on the target model's hidden states.
50%UnverifiedD3D is a standalone diffusion drafter whose core design combines block-level denoising (generating an entire block of candidate tokens in parallel) with auxiliary autoregressive supervision (maintaining a longer high-quality accepted prefix).
50%UnverifiedD3D abandons the use of the target model's hidden states as a conditioning signal and instead compensates drafting quality with autoregressive supervision.
50%UnverifiedD3D can draft subsequent blocks via a single packed pass while the target model performs verification, enabling asynchronous speculative decoding.
50%UnverifiedD3D employs layer-wise progressive pruning to reduce the extra computational overhead of asynchronous packed drafting, with each layer pruned according to its impact on acceptance length.
50%UnverifiedOn three benchmarks with Qwen3-8B as the target model, D3D still retains over 99% of the average acceptance length after pruning 5 of 36 layers.
50%UnverifiedIn evaluations across seven benchmarks, D3D achieves an average decoding speedup of 5.00x with Qwen3-8B as the target model.
50%UnverifiedOn the GSM8K benchmark, D3D reaches a speedup of 7.48x with Qwen3-32B as the target model.
50%UnverifiedD3D's code is open-sourced on GitHub at the repository path fudan-generative-vision/D3D.
50%