# Fast-dLLM [![Project](https://img.shields.io/static/v1?label=Project&message=Github&color=blue&logo=github-pages)](https://nvlabs.github.io/Fast-dLLM) [![arXiv v1](https://img.shields.io/badge/Paper-v1-red.svg)](https://arxiv.org/abs/2505.22618) [![arXiv v2](https://img.shields.io/badge/Paper-v2-red.svg)](https://arxiv.org/abs/2509.26328) [![arXiv dVLM](https://img.shields.io/badge/Paper-dVLM-red.svg)](https://arxiv.org/abs/2604.06832) [![arXiv dDrive](https://img.shields.io/badge/Paper-dDrive-red.svg)](https://arxiv.org/abs/2605.23163)  

ICLR 2026

Fast-dLLM is a family of acceleration techniques for diffusion-based Large Language Models (dLLMs), Vision-Language Models (dVLMs), and Vision-Language-Action (VLA) models. This repository contains: | | Fast-dLLM v1 | Fast-dLLM v2 | Fast-dVLM | Fast-dDrive | |---|---|---|---|---| | **Paper** | [Training-free Acceleration of Diffusion LLM](https://arxiv.org/abs/2505.22618) | [Efficient Block-Diffusion LLM](https://arxiv.org/abs/2509.26328) | [Block-Diffusion VLM via Direct Conversion](https://arxiv.org/abs/2604.06832) | [Efficient Block-Diffusion VLM for Autonomous Driving](https://arxiv.org/abs/2605.23163) | | **Modality** | Text | Text | Vision + Text | Vision + Text + Action (driving) | | **Approach** | Training-free inference acceleration | Block diffusion with fine-tuning | Direct AR-to-diffusion VLM conversion | Section-aware block diffusion + scaffold speculative decoding | | **Backbone** | [Dream](https://github.com/dream-project/dream), [LLaDA](https://github.com/llada-project/llada) | [Qwen2.5](https://github.com/QwenLM/Qwen2.5) | [Qwen2.5-VL](https://github.com/QwenLM/Qwen2.5-VL) | [Qwen2.5-VL](https://github.com/QwenLM/Qwen2.5-VL) | | **Key Techniques** | KV Cache + Parallel Decoding | Block Diffusion + Hierarchical Caching | Block-Size Annealing + Speculative Decoding | SASD Training + Scaffold Spec + Test-Time Inference Scaling | | **Code** | [`v1/`](v1/) | [`v2/`](v2/) | [`fast_dvlm/`](fast_dvlm/) | [`fast_ddrive/`](fast_ddrive/) | | **Model** | β€” | [Fast_dLLM_v2_7B](https://huggingface.co/Efficient-Large-Model/Fast_dLLM_v2_7B) | [Fast_dVLM_3B](https://huggingface.co/Efficient-Large-Model/Fast_dVLM_3B) | [Fast-dDrive](https://huggingface.co/Efficient-Large-Model/Fast-dDrive) | ## News * (πŸ”₯ New) [2026/05/26] **Fast-dDrive** is released! Section-Aware Structured Diffusion VLA for end-to-end autonomous driving on Waymo (WOD-E2E). Combines Scaffold Speculative Decoding with SASD training for SOTA ADE / RFS at over 200 TPS on a single H100 (up to **12x** over the AR baseline with SGLang). Check out [`fast_ddrive/`](fast_ddrive/), the [model](https://huggingface.co/Efficient-Large-Model/Fast-dDrive), and the [paper](https://arxiv.org/abs/2605.23163). * [2026/04/10] **Fast-dVLM** is released! Up to **6.18x speedup** over AR baseline while matching quality across 11 benchmarks. Check out our [webpage](https://nvlabs.github.io/Fast-dLLM/fast_dvlm/), [model](https://huggingface.co/Efficient-Large-Model/Fast_dVLM_3B), and [paper](https://arxiv.org/abs/2604.06832)! * (πŸ”₯ New) [2026/01/26] **Fast-dLLM v1/v2 is accepted by ICLR-2026.** πŸŽ‰πŸŽ‰πŸŽ‰ * \[2025.10.08\] We have open sourced Fast-dLLM v2. Have a look at our [webpage](https://nvlabs.github.io/Fast-dLLM/v2/), [model](https://huggingface.co/Efficient-Large-Model/Fast_dLLM_v2_7B), and [paper](https://arxiv.org/pdf/2509.26328)! * \[2025.08.01\] Our new online demo of Fast-dLLM: https://fast-dllm.hanlab.ai/, welcome to try! * \[2025.07.06\] Added factor-based parallel strategy and LLaDA-1.5 evaluation in `v1/llada/eval_gsm8k.sh`. * \[2025.07.04\] We updated our paper with latest improvements and evaluation results. * \[2025.06.30\] Fast-dLLM has been integrated into [LLaDA-V](https://github.com/ML-GSAI/LLaDA-V). With Fast-dLLM, it accelerates the inference latency from 60s to 6s! Have a try [here](https://github.com/ML-GSAI/LLaDA-V/blob/main/train/generate_demo.py)!! ## TODOs - \[βœ…\] Inference and evaluation code - \[βœ…\] Training code of Fast-dLLM v2 - \[βœ…\] Fast-dVLM: Block-diffusion VLM - \[βœ…\] Fast-dDrive: Block-diffusion VLA for autonomous driving - \[πŸš€\] vLLM support ## Project Structure ``` Fast-dLLM/ β”œβ”€β”€ v1/ # Fast-dLLM v1: Training-free acceleration (LLM) β”‚ β”œβ”€β”€ dream/ # Dream model support β”‚ β”œβ”€β”€ llada/ # LLaDA model support β”‚ β”œβ”€β”€ requirements.txt β”‚ └── README.md β”œβ”€β”€ v2/ # Fast-dLLM v2: Block diffusion (LLM) β”‚ β”œβ”€β”€ src/ # LMFlow training framework β”‚ β”œβ”€β”€ train_scripts/ # Fine-tuning scripts β”‚ β”œβ”€β”€ configs/ # DeepSpeed configs β”‚ β”œβ”€β”€ generation_functions.py β”‚ β”œβ”€β”€ eval.py / eval_script.sh β”‚ β”œβ”€β”€ app.py / run_chatbot.py β”‚ β”œβ”€β”€ requirements.txt β”‚ └── README.md β”œβ”€β”€ fast_dvlm/ # Fast-dVLM: Block-diffusion VLM (chatbot, optional finetune sample, VLMEval; see fast_dvlm/README.md) β”œβ”€β”€ fast_ddrive/ # Fast-dDrive: Block-diffusion VLA for autonomous driving on Waymo E2E (see fast_ddrive/README.md) β”œβ”€β”€ CONTRIBUTING.md β”œβ”€β”€ LICENSE └── README.md # This file ``` ## Quick Start ### Fast-dLLM v1 (Training-free Acceleration) ```bash cd v1 pip install -r requirements.txt # LLaDA interactive chat python llada/chat.py --gen_length 128 --steps 128 --block_size 32 # Dream evaluation accelerate launch dream/eval.py --model dream \ --model_args pretrained=Dream-org/Dream-v0-Base-7B,max_new_tokens=256,diffusion_steps=8,add_bos_token=true,alg=confidence_threshold,threshold=0.9,use_cache=true \ --tasks gsm8k --num_fewshot 5 --batch_size 1 ``` For full details, see [v1/README.md](v1/README.md). ### Fast-dLLM v2 (Block Diffusion) ```bash cd v2 pip install -e . # Gradio web demo python app.py # Evaluation bash eval_script.sh ``` For full details, see [v2/README.md](v2/README.md). ### Fast-dVLM (Block-Diffusion VLM) ```bash cd fast_dvlm pip install -r requirements.txt # Quick inference python run_chatbot.py \ --model-name Efficient-Large-Model/Fast_dVLM_3B \ --image path/to/image.jpg \ --prompt "Describe this image in detail." # Interactive mode python run_chatbot.py ``` **Fine-tuning (optional example):** multimodal MDM training uses DeepSpeed + the LMFlow fork under [`third_party/`](third_party/) (the launcher sets `PYTHONPATH` for you). Download [ALLaVA-4V](https://huggingface.co/datasets/FreedomIntelligence/ALLaVA-4V) with `fast_dvlm/data/download_example_dataset.sh`, then run `bash fast_dvlm/train_scripts/finetune_multimodal_example.sh` from the repo rootβ€”see [Fine-tuning (example launcher)](fast_dvlm/README.md#fine-tuning-example-launcher) in [fast_dvlm/README.md](fast_dvlm/README.md). For full details, see [fast_dvlm/README.md](fast_dvlm/README.md). ### Fast-dDrive (Block-Diffusion VLA for Autonomous Driving) ```bash cd fast_ddrive pip install -r requirements.txt # Single-shot demo: Scaffold Spec decoding on one driving frame. python run_chatbot.py \ --model_path Efficient-Large-Model/Fast-dDrive \ --image data/example/images/161_CAM_FRONT.jpg \ --prompt "Describe the driving scene and produce a 5-second plan." # Waymo E2E validation eval (paper canonical Scaffold Spec, multi-GPU). MODEL_PATH=Efficient-Large-Model/Fast-dDrive EVAL_JSON=/path/to/waymo_val.json \ IMAGE_ROOT=/path/to/image_root bash run_eval.sh ``` Three decoding paths are exposed via `--mode` / `MODE`: `section_diffusion` (SD), `scaffold_spec` (SS β€” paper canonical), and `inference_scaling` (SS multi-trajectory rollouts). **Fine-tuning (SASD):** mirrors the fast_dvlm DeepSpeed launcher and reuses the same vendored LMFlow under [`third_party/`](third_party/) (with a small set of pure-addition SASD hooks). Provide a Waymo training JSON + image root, then: ```bash DATASET_PATH=/path/to/waymo_train.json IMAGE_FOLDER=/path/to/image_root \ bash fast_ddrive/train_scripts/train_waymo_sasd.sh ``` For full details, see [fast_ddrive/README.md](fast_ddrive/README.md). ## Contributing Issues and Pull Requests are welcome! Please see [CONTRIBUTING.md](CONTRIBUTING.md) for details. ## License This project is licensed under the Apache License 2.0. See the [LICENSE](LICENSE) file for details. ## Citation If you find this work useful, please cite our papers: ```bibtex @misc{zhang2026fastddriveefficientblockdiffusionvlm, title={Fast-dDrive: Efficient Block-Diffusion VLM for Autonomous Driving}, author={Kewei Zhang and Jin Wang and Sensen Gao and Chengyue Wu and Yulong Cao and Songyang Han and Boris Ivanovic and Langechuan Liu and Marco Pavone and Song Han and Daquan Zhou and Enze Xie}, year={2026}, eprint={2605.23163}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2605.23163}, } @misc{wu2026fastdvlmefficientblockdiffusionvlm, title={Fast-dVLM: Efficient Block-Diffusion VLM via Direct Conversion from Autoregressive VLM}, author={Chengyue Wu and Shiyi Lan and Yonggan Fu and Sensen Gao and Jin Wang and Jincheng Yu and Jose M. Alvarez and Pavlo Molchanov and Ping Luo and Song Han and Ligeng Zhu and Enze Xie}, year={2026}, eprint={2604.06832}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2604.06832}, } @misc{wu2025fastdllmv2efficientblockdiffusion, title={Fast-dLLM v2: Efficient Block-Diffusion LLM}, author={Chengyue Wu and Hao Zhang and Shuchen Xue and Shizhe Diao and Yonggan Fu and Zhijian Liu and Pavlo Molchanov and Ping Luo and Song Han and Enze Xie}, year={2025}, eprint={2509.26328}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2509.26328}, } @misc{wu2025fastdllmtrainingfreeaccelerationdiffusion, title={Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decoding}, author={Chengyue Wu and Hao Zhang and Shuchen Xue and Zhijian Liu and Shizhe Diao and Ligeng Zhu and Ping Luo and Song Han and Enze Xie}, year={2025}, eprint={2505.22618}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2505.22618}, } ``` ## Acknowledgements We would like to thank the authors of [LLaDA](https://github.com/llada-project/llada) and [Dream](https://github.com/dream-project/dream) for their excellent work and open-source contributions. We thank [Qwen2.5](https://github.com/QwenLM/Qwen2.5) and [Qwen2.5-VL](https://github.com/QwenLM/Qwen2.5-VL) for the base model architectures and [LMFlow](https://github.com/OptimalScale/LMFlow) for the training framework.