d3LLM: Ultra-Fast Diffusion LLM using Pseudo-Trajectory Distillation 🚀
This repository contains d3LLM-LLaDA, an ultra-fast diffusion language model presented in the paper d3LLM: Ultra-Fast Diffusion LLM using Pseudo-Trajectory Distillation.
- 📄 Paper: arXiv:2601.07568
- 💻 Code: GitHub - hao-ai-lab/d3LLM
- 🌐 Blog: Ultra-Fast Diffusion LLMs
- 🕹️ Demo: d3LLM Demo
Model Description
d3LLM-LLaDA is an ultra-fast diffusion language model that strikes a balance between accuracy and parallelism. It uses pseudo-trajectory distillation to teach the model which tokens can be decoded confidently at early steps, and employs an entropy-based multi-block decoding mechanism with KV-cache refresh during inference.
Key Features
- 🚀 High throughput: 5.0× faster than autoregressive models (Qwen-2.5-7B-it) on H100 GPU and 3.5× faster on A100 GPU.
- 📊 High AUP: Achieves high Accuracy Under Parallelism scores across benchmarks.
- 🔧 Task Optimization: Specifically optimized for coding and math reasoning tasks.
Installation
To use this model, it is recommended to clone the official repository and install the required dependencies:
# Clone the repository
git clone https://github.com/hao-ai-lab/d3LLM.git
cd d3LLM
# Install dependencies
pip install -r requirements.txt
Citation
If you find d3LLM useful for your research, please cite the following work:
@inproceedings{ICML'26:d3llm,
title = {d3LLM: Ultra-Fast Diffusion LLM using Pseudo-Trajectory Distillation},
author = {Yu-Yang Qian and Junda Su and Lanxiang Hu and Peiyuan Zhang and Zhijie Deng and Peng Zhao and Hao Zhang},
booktitle = {Proceedings of the 43rd International Conference on Machine Learning (ICML)},
pages = {to appear},
year = {2026}
}