DDTree

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

🌐 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/.

Reproduce Paper Artifacts

Generate the plots:

python3 plot_results.py

Generate the LaTeX table:

python3 make_latex_table.py