DDTree

Official implementation of DDTree (Diffusion Draft Tree) from Accelerating Speculative Decoding with Block Diffusion Draft Trees.

Liran Ringel, Yaniv Romano

🌐 Project Page  |  📄 Paper

Setup

This codebase is intended for a CUDA-enabled PyTorch environment.

pip install -r requirements.txt

Run Experiments

bash run_benchmark.sh

This produces benchmark outputs in runs/ and logs in logs/.

Single-GPU Targeted Run (Recommended for Local Debugging)

To run one model on one dataset on one GPU:

CUDA_VISIBLE_DEVICES=0 \
NPROC_PER_NODE=1 \
TASKS_OVERRIDE="gsm8k:128" \
MODEL_DRAFT_PAIRS_OVERRIDE="Qwen/Qwen3-4B|z-lab/Qwen3-4B-DFlash-b16" \
TEMPERATURES_OVERRIDE="0.0" \
MODES_OVERRIDE="sdpa" \
DISABLE_CPP_COMPACT_CACHE=1 \
bash run_benchmark.sh

Useful overrides:

• MAX_SAMPLES_OVERRIDE (for quick smoke tests)
• MAX_NEW_TOKENS (default 2048)
• TREE_BUDGET_OVERRIDE (for DDTree budget list)
• MODES_OVERRIDE="sdpa" to avoid running both modes

Reproduce Paper Artifacts

Generate the plots:

python3 plot_results.py

Generate the LaTeX table:

python3 make_latex_table.py

Citation

@article{ringel2026ddtree,
  title={Accelerating Speculative Decoding with Block Diffusion Draft Trees},
  author={Ringel, Liran and Romano, Yaniv},
  journal={arXiv preprint arXiv:2604.12989},
  year={2026}
}