File size: 2,306 Bytes
ae3482e 5098454 8796070 ae3482e 8796070 ae3482e 8796070 ae3482e 8796070 5098454 8796070 ae3482e 8796070 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 |
---
datasets:
- d3LLM/Ling-Coder-dParallel-merged-512-120k
base_model: Dream-org/Dream-Coder-v0-Instruct-7B
pipeline_tag: text-generation
library_name: transformers
license: apache-2.0
tags:
- diffusion
- fast-inference
- d3llm
---
# d3LLM: Ultra-Fast Diffusion LLM using Pseudo-Trajectory Distillation 🚀
**d3LLM-Dream-Coder** is an ultra-fast diffusion language model introduced in the paper [d3LLM: Ultra-Fast Diffusion LLM using Pseudo-Trajectory Distillation](https://huggingface.co/papers/2601.07568). It is built on [Dream-org/Dream-Coder-v0-Instruct-7B](https://huggingface.co/Dream-org/Dream-Coder-v0-Instruct-7B).
## Model Description
d3LLM (pseuDo-Distilled Diffusion Large Language Model) is a framework designed to strike a balance between accuracy and parallelism in diffusion LLMs. It achieves up to 10× speedup over vanilla diffusion models like LLaDA/Dream and 5× speedup over autoregressive (AR) models.
The model utilizes two primary innovations:
- **Pseudo-Trajectory Distillation**: A training method that teaches the model which tokens can be decoded confidently at early steps.
- **Entropy-Based Multi-Block Decoding**: An inference strategy using a KV-cache refresh mechanism to maintain accuracy while maximizing parallelism.
## Resources
- **Paper**: [d3LLM: Ultra-Fast Diffusion LLM using Pseudo-Trajectory Distillation](https://huggingface.co/papers/2601.07568)
- **Repository**: [https://github.com/hao-ai-lab/d3LLM](https://github.com/hao-ai-lab/d3LLM)
- **Blog**: [https://hao-ai-lab.github.io/blogs/text-diffusion/](https://hao-ai-lab.github.io/blogs/text-diffusion/)
- **Demo**: [https://d3llm-team.github.io/](https://d3llm-team.github.io/)
## Usage
For detailed usage instructions, evaluation scripts, and training code, please refer to the official GitHub repository. Since the model uses a custom architecture, ensure you have `transformers==4.49.0` installed and use `trust_remote_code=True` when loading the model.
## Citation
```bibtex
@article{arxiv'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},
journal = {ArXiv preprint},
volume = {arXiv:2601.07568},
year = {2026}
}
``` |