| --- |
| license: mit |
| 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-2: Target-Aligned Parallel Draft Model for Dual-Mode Speculative Decoding</h1> |
| </div> |
|
|
|
|
| <p align="center"> | |
| <a href="https://arxiv.org/pdf/2605.08632"><b>Paper</b></a> | |
| <a href="https://github.com/AMD-AIG-AIMA/PARD"><b>Github</b></a> | |
| </p> |
|
|
| <br clear="left"/> |
|
|
| ## Introduction |
|
|
| PARD is a high-performance speculative decoding method that also enables low-cost adaptation of autoregressive draft models into parallel draft models. PARD-2 further advances PARD by introducing a Target-Aligned Parallel Draft Model for dual-mode speculative decoding. Instead of optimizing draft models only for token-level prediction accuracy, PARD-2 aligns draft-model training with the inference-time objective of maximizing consecutive token acceptance. PARD-2 offers the following advantages: |
|
|
| - **Target-Aligned Optimization**: PARD-2 reformulates the draft-model objective from next-token prediction accuracy to acceptance-length optimization, better matching the draft-then-verify process used during speculative decoding. |
|
|
| - **Confidence-Adaptive Token Optimization**: PARD-2 introduces Confidence-Adaptive Token (CAT) optimization, which adaptively reweights tokens according to their contribution to the verification process. This improves the alignment between draft generation and target-model acceptance. |
|
|
| - **Dual-Mode Speculative Decoding**: A single PARD-2 draft model supports both target-independent and target-dependent modes, combining the deployment flexibility of PARD with the stronger alignment capability of target-aware methods. |
|
|
| State-of-the-Art Performance: Across diverse models and tasks, PARD-2 achieves up to 6.94× lossless acceleration. On LLaMA3.1-8B, PARD-2 surpasses EAGLE-3 by 1.9× and PARD by 1.3×, setting a new performance frontier for speculative decoding. |
|
|
|
|
| <p align="center"> |
| <img src="https://raw.githubusercontent.com/AMD-AGI/PARD/master/datas/img/pard_2.png" width="100%"> |
| <br> |
| <em>Throughput and Latency Trade-offs on vLLM. PARD-2 consistently achieves a superior Pareto frontier across various batch sizes from 1 to 64.</em> |
| </p> |
|
|
|
|
| ## Model Weights |
|
|
| | Model Series | Model Name | Download | |
| |--------------|---------------------------------------|---------------| |
| | Llama3 | amd/PARD2-Llama-3.1-8B | [🤗 HuggingFace](https://huggingface.co/amd/PARD2-Llama-3.1-8B ) | |
| | Qwen3 | amd/PARD2-Qwen3-8B | [🤗 HuggingFace](https://huggingface.co/amd/PARD2-Qwen3-8B) | |
| | Qwen3 | amd/PARD2-Qwen3-14B | [🤗 HuggingFace](https://huggingface.co/amd/PARD2-Qwen3-14B) | |
|
|
| ## How To Use |
|
|
| Please visit [PARD2](https://github.com/AMD-AIG-AIMA/PARD) repo for more information |
|
|
|
|
| ## Citation |
| ``` |
| @article{an2026pard, |
| title={PARD-2: Target-Aligned Parallel Draft Model for Dual-Mode Speculative Decoding}, |
| author={An, Zihao and Liu, Taichi and Liu, Ziqiong and Li, Dong and Liu, Ruofeng and Barsoum, Emad}, |
| journal={arXiv preprint arXiv:2605.08632}, |
| year={2026} |
| } |
| ``` |
