|
|
---
|
|
|
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}
|
|
|
}
|
|
|
```
|
|
|
|
