| # Fast-dLLM v1: Training-free Acceleration of Diffusion LLM |
|
|
| [](https://nvlabs.github.io/Fast-dLLM) |
| [](https://arxiv.org/abs/2505.22618) |
| <a href="https://fast-dllm.hanlab.ai"><img src="https://img.shields.io/static/v1?label=Demo&message=Fast-dLLM&color=yellow"></a>   |
|
|
| Fast-dLLM v1 is a **training-free** inference acceleration framework for diffusion-based Large Language Models (dLLMs). It supports efficient inference for models like **Dream** and **LLaDA** by enabling KV Cache and Parallel Decoding. |
|
|
| ## Key Features |
|
|
| 1. **Key-Value Cache for Block-Wise Decoding** |
| We propose an efficient block-wise decoding KV Cache mechanism for Masked Diffusion Models (MDMs). By reusing attention Key-Value activations across multiple steps within each block, our approach avoids redundant computation and significantly accelerates inference. Furthermore, our DualCache extension also caches masked suffix tokens, enabling even greater speedup with negligible accuracy loss. |
|
|
| <div align="center"> |
| <img src="asset/kvcache.jpg" alt="KV Cache for block-wise decoding" width="800"/> |
| <p>KV Cache for block-wise decoding</p> |
| </div> |
|
|
| 2. **Confidence-Aware Parallel Decoding** |
| Instead of decoding tokens sequentially, we introduce a confidence-aware parallel decoding scheme. At each step, only tokens with confidence over a threshold are unmasked in parallel, while uncertain ones remain masked for future steps. This selective approach effectively balances decoding efficiency and output quality. |
|
|
| <div align="center"> |
| <img src="asset/output.gif" alt="Decoding comparison" width="800"/> |
| <p><b>Left:</b> Standard decoding (LLaDA). <b>Right:</b> Confidence-aware parallel decoding.</p> |
| </div> |
|
|
| <div align="center"> |
| <img src="asset/pseudo_code.jpg" alt="Pseudo code for our method" width="800"/> |
| <p>Pseudo code for our method</p> |
| </div> |
|
|
| 3. **Overall Performance** |
| Overall, introducing the KV Cache mechanism yields significant speed improvements for all tasks and sequence lengths, typically achieving a 2x to 3.6x speedup compared to the vanilla backbone. When the parallel decoding strategy is applied individually, we see additional acceleration, often pushing speedups to 4x-6x for the evaluated settings, particularly as the generation length increases. |
|
|
| <div align="center"> |
| <img src="asset/overall_performance.jpg" alt="Overall performance" width="800"/> |
| <p>Overall performance comparison</p> |
| </div> |
|
|
| ## Demo |
|
|
| https://github.com/user-attachments/assets/32bbff97-6e60-4e14-95c0-2cbec136476f |
|
|
| <div align="center"> |
| <img src="asset/speedup.jpg" alt="End-to-end speedup over vanilla LLaDA baseline" width="800"/> |
| <p>End-to-end speedup over vanilla LLaDA baseline</p> |
| </div> |
|
|
| ## File Structure |
|
|
| ``` |
| v1/ |
| βββ README.md # This file |
| βββ requirements.txt # Dependencies for inference & evaluation |
| βββ dream/ # Dream model related code |
| β βββ model/ # Dream model definition |
| β βββ eval.py # Evaluation harness integration |
| β βββ eval.md # Evaluation guide |
| β βββ eval_gsm8k.sh # GSM8K evaluation script |
| β βββ eval_humaneval.sh # HumanEval evaluation script |
| β βββ demo_multiturn_chat.py # Multi-turn chat demo |
| βββ llada/ # LLaDA model related code |
| βββ model/ # LLaDA model definition |
| βββ generate.py # Core generation with cache & parallel decoding |
| βββ eval_llada.py # Evaluation harness integration |
| βββ eval.md # Evaluation guide |
| βββ eval_gsm8k.sh # GSM8K evaluation script |
| βββ eval_humaneval.sh # HumanEval evaluation script |
| βββ chat.py # Command-line chat interface |
| βββ app.py # Gradio web demo |
| ``` |
|
|
| ## Installation |
|
|
| ```bash |
| cd v1 |
| pip install -r requirements.txt |
| ``` |
|
|
| ## Usage |
|
|
| ### 1. Using LLaDA Model |
|
|
| #### Interactive Chat |
| ```bash |
| python llada/chat.py --gen_length 128 --steps 128 --block_size 32 |
| ``` |
|
|
| Parameter descriptions: |
| - `--gen_length`: Maximum length of generated text |
| - `--steps`: Number of sampling steps |
| - `--block_size`: Cache block size |
| - `--use_cache`: Whether to use cache |
| - `--if_cache_position`: Whether to use dual cache |
| - `--threshold`: Confidence threshold |
|
|
| #### Web Demo |
| ```bash |
| pip install gradio |
| cd llada |
| python app.py |
| ``` |
|
|
| #### Model Evaluation |
| | Benchmark | Gen Length | LLaDA | +Cache | +Parallel | +Cache+Parallel (Fast-dLLM) | |
| |-------------------|------------|---------|----------------|----------------|-----------------------------| |
| | **GSM8K (5-shot)**| 256 | 79.3<br>6.73<br>(1Γ) | 79.5<br>21.23<br>(3.2Γ) | 79.2<br>16.53<br>(2.5Γ) | 78.5<br>**54.4<br>(8.1Γ)** | |
| | | 512 | 77.5<br>3.23<br>(1Γ) | 77.0<br>10.43<br>(3.3Γ) | 77.6<br>18.63<br>(5.8Γ) | 77.2<br>**35.3<br>(11.0Γ)** | |
| | **HumanEval (0-shot)** | 256 | 41.5<br>30.5 (1Γ) | 42.7<br>40.73<br>(1.3Γ) | 43.9<br>101.53<br>(3.3Γ) | 43.3<br>**114.1<br>(3.7Γ)** | |
| | | 512 | 43.9<br>18.4 (1Γ) | 45.7<br>29.33<br>(1.6Γ) | 43.3<br>57.13<br>(3.1Γ) | 44.5<br>**73.7<br>(4.0Γ)** | |
|
|
| Each cell presents the accuracy (top row, in percentage) and the decoding throughput (middle row, in tokens per second) with relative speedup (bottom row) to the LLaDA baseline. |
|
|
| For detailed evaluation instructions, please refer to: |
| - [LLaDA Evaluation Guide](llada/eval.md) |
| - [Dream Evaluation Guide](dream/eval.md) |
|
|
| ### 2. Using Dream Model |
|
|
| For detailed evaluation instructions on GSM8K and HumanEval benchmarks, please refer to [Dream Evaluation Guide](dream/eval.md). |
|
|
| ## Citation |
|
|
| ```bibtex |
| @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. |
|
|