---
license: mit
language:
- en
library_name: transformers
pipeline_tag: text-generation
base_model:
- GSAI-ML/LLaDA-8B-Instruct
tags:
- diffusion-language-model
- dllm
- llada
---

# LLaDA-8B-Instruct-HF

This repository hosts a re-formatted version of [GSAI-ML/LLaDA-8B-Instruct](https://huggingface.co/GSAI-ML/LLaDA-8B-Instruct), a large language **diffusion** model (LLaDA). Specifically, we convert the original LLaDA weights into the **LLM-part format used by LLaDA-V** (the LLaVA-style multimodal LLaDA), so the checkpoint can be directly plugged in as the language backbone of our [PerceptionDLM](https://github.com/MSALab-PKU/PerceptionDLM) training and inference pipeline.

The weights are functionally identical to the original LLaDA-8B-Instruct; only the parameter-key naming and index/config files are adapted to match the LLaDA-V LLM module. It is used as the **diffusion language backbone** for [PerceptionDLM-Base](https://huggingface.co/MSALab/PerceptionDLM-Base) and [PerceptionDLM](https://huggingface.co/MSALab/PerceptionDLM).

<p align="center">
  📄 <a href="https://arxiv.org/abs/2606.19534">PerceptionDLM Paper</a> &nbsp;|&nbsp;
  💻 <a href="https://github.com/MSALab-PKU/PerceptionDLM">Code</a> &nbsp;|&nbsp;
  🔗 <a href="https://huggingface.co/GSAI-ML/LLaDA-8B-Instruct">Original LLaDA-8B-Instruct</a>
</p>

## What is different from the original?

- ✅ Same model weights and behavior as `GSAI-ML/LLaDA-8B-Instruct`.
- 🔧 Weights re-formatted from the original LLaDA layout into the **LLaDA-V LLM-part format** (parameter-key renaming + regenerated safetensors index/config).
- 📦 Ready to be used directly as `language_model.name_or_path` in our training configs.

You can also reproduce this conversion yourself:

```bash
python scripts/convert.py \
  --model_path /path/to/LLaDA-8B-Instruct \
  --output /path/to/LLaDA-8B-Instruct-HF
```

## Usage

This checkpoint is primarily intended to be consumed by the PerceptionDLM pipeline. Please refer to the [GitHub repository](https://github.com/MSALab-PKU/PerceptionDLM) for setup and usage. For standalone LLaDA usage and generation details, see the [original LLaDA repository](https://github.com/ML-GSAI/LLaDA).

## Citation

If you use this checkpoint, please cite both LLaDA and PerceptionDLM:

```bibtex
@article{nie2025llada,
  title   = {Large Language Diffusion Models},
  author  = {Nie, Shen and Zhu, Fengqi and You, Zebin and Zhang, Xiaolu and Ou, Jingyang and Hu, Jun and Zhou, Jun and Lin, Yankai and Wen, Ji-Rong and Li, Chongxuan},
  journal = {arXiv preprint arXiv:2502.09992},
  year    = {2025}
}

@article{sun2026perceptiondlm,
  title   = {PerceptionDLM: Parallel Region Perception with Multimodal Diffusion Language Models},
  author  = {Sun, Yueyi and Wang, Yuhao and Li, Jason and Tian, Ye and Zhang, Tao and Mai, Jacky and Wang, Yihan and Wang, Haochen and Bai, Jinbin and Yang, Ling and Tong, Yunhai},
  journal = {arXiv preprint arXiv:2606.19534},
  year    = {2026}
}
```

## License

This conversion follows the license of the original **GSAI-ML/LLaDA-8B-Instruct** release. Please consult the [original model card](https://huggingface.co/GSAI-ML/LLaDA-8B-Instruct) for the authoritative license terms before use.