Speculative Decoding Meets Quantization: Compatibility Evaluation and Hierarchical Framework Design

Introduction

We present a systematic evaluation of the compatibility between speculative decoding and quantization.

We also propose a hierarchical speculative decoding framework for W4A16 models, achieving a 1.31$\times$ speedup over EAGLE-2.

And all experiments are implemented in C/CUDA.

Speedup

Speedup achieved by integrating Speculative Decoding and Quantization.

Installation from source

conda create -n specmquant python=3.11 && conda activate specmquant
# install pytorch for your platform, see https://pytorch.org
git clone https://github.com/AI9Stars/SpecMQuant --recursive && cd SpecMQuant
vim setup.py # change arch="80" to other code for your platform, see https://developer.nvidia.com/cuda-gpus#compute
pip install .

Evaluation

Model Preparation

Downloads the quantized model weights and corresponding EAGLE model into the models folder.

---

Or you can also choose one of the following external toolkits to quantize your model and then convert the resulting checkpoints.

1. Supported Toolkits & Precision

For AutoGPTQ, our framework is only compatible when sym=True is set in the config, and if you set desc_act=True then you should also set static_group=True.

2. Model Convert

For W4A16, W4A8-QQQ, W4A8-QQQ-g128 and W4A8-QoQ-g128, after quantizing with the above toolkits you need to convert the model checkpoints using the scripts in scripts/model_convert. And for the models applied with rotation method, you need to convert the eagle checkpoint using the scripts scripts/model_convert/convert_eagle_rotation.sh with the corresponding rotation matrix.

---

Run Evaluation

MT-Bench

All scripts for MT-Bench evaluation are located in the scripts/eval/mt_bench folder. Here we use Llama-3-8B-Instruct as an example:

# 1. Run evaluations
bash scripts/eval/mt_bench/llama3-8b-instruct/<precision>/run_baseline.sh
bash scripts/eval/mt_bench/llama3-8b-instruct/<precision>/run_eagle.sh

# 2. Evaluate speed
bash scripts/mt_bench/llama3-8b-instruct/speed_up.sh

Replace <precision> with one of: fp16, w4a16, w4a8-qqq, w4a8-qqq-g128, w4a8-qoq, or w4a8-qoq-g128.

---

Spec-Bench

Scripts for Spec-Bench evaluation in W4A16 Llama-3-70B-Instruct models are located in the scripts/eval/spec_bench folder.

# 1. Run evaluations
bash scripts/eval/spec_bench/llama3-70b-instruct-w4a16/run_baseline.sh
bash scripts/eval/spec_bench/llama3-70b-instruct-w4a16/run_spec.sh
bash scripts/eval/spec_bench/llama3-70b-instruct-w4a16/run_eagle.sh
bash scripts/eval/spec_bench/llama3-70b-instruct-w4a16/run_hierspec.sh


# 2. Evaluate speed
bash scripts/eval/spec_bench/llama3-70b-instruct-w4a16/speedup.sh

---

Performance evaluation

We provide the performance evaluation for gsm8k and human_eval.

# 1. Run evaluations
bash scripts/eval/<benchmark>/llama3-8b-instruct/<precision>/run_baseline.sh

# 2. Evaluate preformance
bash scripts/eval/<benchmark>/llama3-8b-instruct/check_correctness.sh

Replace <benchmark> with one of: gsm8k or human_eval.

Contributors

• Yudi Zhang
• Weilin Zhao

Acknowledgment

Our framework is based on https://github.com/thunlp/FR-Spec.

Our experiments are based on https://github.com/SafeAILab/EAGLE.

The CUDA quantization kernels in src/qgemm are borrowed from:

• W4A16 marlin kernel:https://github.com/vllm-project/vllm and https://github.com/IST-DASLab/marlin.
• W4A8-QQQ kernel: https://github.com/HandH1998/QQQ.
• W8A8 and W4A8-QoQ: https://github.com/mit-han-lab/omniserve.

The evaluation/ folder is modified base on https://github.com/hemingkx/Spec-Bench:

• The evaluation/gsm8k folder integrates part of the code from https://github.com/Guangxuan-Xiao/GSM8K-eval.
• The evaluation/humaneval folder integrates part of the code from https://github.com/evalplus/evalplus.

The src/flash_attn/ folder is modified base on https://github.com/Dao-AILab/flash-attention/blob/v2.4.2/csrc/flash_attn.

Citation

@article{zhang2025specmqaunt,
  title={Speculative Decoding Meets Quantization: Compatibility Evaluation and Hierarchical Framework Design},
  author={Zhang, Yudi and Zhao, Weilin and Han, Xu and Zhao, Tiejun and Xu, Wang and Cao, Hailong and Zhu, Conghui},
  journal={arXiv preprint arXiv:2505.22179},
  year={2025}
}