Official implementation of DDTree (Diffusion Draft Tree) from Accelerating Speculative Decoding with Block Diffusion Draft Trees.
Liran Ringel, Yaniv Romano
π Project Page | π Paper
This codebase is intended for a CUDA-enabled PyTorch environment.
pip install -r requirements.txt
bash run_benchmark.sh
This produces benchmark outputs in runs/ and logs in logs/.
Generate the plots:
python3 plot_results.py
Generate the LaTeX table:
python3 make_latex_table.py
@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} }