D3D-8B-32B
8B draft for Qwen3-32B, released for D3D: Decoupled Diffusion Drafting for Asynchronous Speculative Decoding.
Code & quick start · All checkpoints
Model details
| Property | Value |
|---|---|
| Role | Diffusion draft model; not the target verifier |
| Target verifier | Qwen/Qwen3-32B |
| Draft base architecture | Qwen3-8B |
| Transformer layers | 36 |
| Hidden size | 4096 |
| Vocabulary size | 151936 |
| Stored weight dtype | FP32 (original checkpoint tensors, no conversion) |
| Recommended inference dtype | BF16 |
Unpruned Qwen3-derived draft architecture.
The size in the model name describes the base draft family. The stored checkpoint
can include untied embedding/output weights, and pruned variants have fewer layers;
the exact tensor parameter count is reported in release_metadata.json.
Quick start
Install the dependencies in the D3D repository, then run from its root:
export TARGET_MODEL=Qwen/Qwen3-32B
export DRAFT_MODEL=fudan-generative-ai/D3D-8B-32B
bash scripts/quickstart_single.sh
For concurrent requests:
CONCURRENCY=4 bash scripts/quickstart_concurrent.sh
The target runs on GPU 0 and the draft on GPU 1. D3D uses its checked-in D3D compatibility classes and does not execute Hub-hosted Python during inference. Use the pinned D3D environment (PyTorch 2.8.0 / Transformers 4.57.3).