fudan-generative-ai/D3D-1.7B-8B
Text Generation • 2B • Updated • 390 • 2
Decoupled Diffusion Drafting for Asynchronous Speculative Decoding. Six Qwen3-derived diffusion draft checkpoints for Qwen3-8B and Qwen3-32B.
Note Qwen3-8B · 1.7B-family draft · 28 layers
Note Qwen3-32B · 1.7B-family draft · 28 layers
Note Qwen3-8B · 8B-family draft · 36 layers
Note Qwen3-32B · 8B-family draft · 36 layers
Note Qwen3-32B · pruned 1.7B-family draft · 24 layers
Note Qwen3-8B · pruned 8B-family draft · 31 layers