File size: 3,449 Bytes
6804ff8
 
 
 
 
 
 
 
 
 
 
0852ada
6804ff8
 
 
 
 
 
 
0852ada
6804ff8
 
 
 
 
 
 
 
 
 
1e62d38
0852ada
1e62d38
7b4425f
1e62d38
 
 
 
0852ada
6804ff8
1e62d38
6804ff8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
---

license: mit
pipeline_tag: text-generation
library_name: transformers
---


<img src="https://cdn-uploads.huggingface.co/production/uploads/66a056d0229269a861ac1245/UmJOD5HnhCfvy3nAXgxgE.png" alt="PARD" width="100" align="left">
<div align="center">
<h1>PARD: Accelerating LLM Inference with Low-Cost PARallel Draft Model Adaptation</h1>
</div>


<p align="center"> |
<a href="https://arxiv.org/abs/2504.18583"><b>Paper</b></a> |
<a href="https://github.com/AMD-AIG-AIMA/PARD"><b>Github</b></a> |
<a href="https://www.amd.com/en/developer/resources/technical-articles/accelerating-generative-llms-interface-with-parallel-draft-model-pard.html"><b>Blog</b></a> |
</p>



## Introduction

PARD is a high-performance speculative decoding method that also enables low-cost adaptation of autoregressive draft models into parallel draft models. It offers the following advantages:

- **Low-Cost Training**: PARD adapts AR (autoregressive) draft models into parallel draft models with minimal overhead. Compared to pure AR draft models, PARD achieves an average inference speedup of 1.78×. By introducing a conditional drop-token strategy, PARD improves training efficiency by up to 3× while maintaining the same level of accuracy.

- **Generalizability**: Thanks to its target-independent design, a single PARD draft model can accelerate an entire family of target models. This contrasts with target-dependent approaches such as Medusa and EAGLE, which require retraining or tuning for each new target. As a result, PARD significantly reduces both deployment complexity and adaptation cost.

- **High Performance**: When integrated into an optimized inference framework called Transformers+ PARD delivers up to a 4.08× speedup, with LLaMA3.1 8B reaches a state-of-the-art 311.5 tokens per second. When integrated into vLLM, PARD delivers up to 3.06× speedup, outperforming other speculative decoding methods in vLLM by 1.51×.


<p align="center">
  <figure style="display: inline-block; text-align: center;">
    <img src="https://cdn-uploads.huggingface.co/production/uploads/630cb01cc169245d78fe76b6/Dh-7wE-l0YAfU9lXWssKf.png" width="100%">

    <figcaption style="font-style: italic; margin-top: 2px;">

      AR and AR+ represent baseline auto-regressive generation using Transformers and Transformers+, respectively. VSD denotes vanilla speculative decoding. PARD refers to the proposed method in this work.

    </figcaption>

  </figure>

</p>



## Model Weights

| Model Series | Model Name                            | Download      |
|--------------|---------------------------------------|---------------|
| llama3       | PARD-Llama-3.2-1B                     | [🤗 HuggingFace](https://huggingface.co/amd/PARD-Llama-3.2-1B)  |
| DSR Qwen     | PARD-DeepSeek-R1-Distill-Qwen-1.5B    | [🤗 HuggingFace](https://huggingface.co/amd/PARD-DeepSeek-R1-Distill-Qwen-1.5B) |
| Qwen         | PARD-Qwen2.5-0.5B                     | [🤗 HuggingFace](https://huggingface.co/amd/PARD-Qwen2.5-0.5B) |


## How To Use

Please visit [PARD](https://github.com/AMD-AIG-AIMA/PARD) repo for more information


## Citation
```

@article{an2025pard,

  title={PARD: Accelerating LLM Inference with Low-Cost PARallel Draft Model Adaptation},

  author={An, Zihao and Bai, Huajun and Liu, Ziqiong and Li, Dong and Barsoum, Emad},

  journal={arXiv preprint arXiv:2504.18583},

  year={2025}

}

```