File size: 9,300 Bytes
dab3c34 1499273 dab3c34 |
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 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 |
---
datasets:
- HuggingFaceTB/smollm-corpus
language:
- en
---
# Outlier-Safe Pre-Training
[](https://arxiv.org/abs/2506.19697)
[](https://huggingface.co/collections/dmis-lab/outlier-safe-pre-training-osp-685bda10aa1e8a19fcb58ea8)
[](https://github.com/dmis-lab/Outlier-Safe-Pre-Training)
## Introduction
Quantization plays a crucial role in deploying Large Language Models (LLMs) in resource-constrained environments. However, the presence of outlier features significantly hinders low-bit quantization. While many studies address this problem in a post-hoc manner to make use of already pre-trained models, the importance of handling outliers during pre-training is often underestimated.
Our work, **Outlier-Safe Pre-Training (OSP)**, proposes a practical approach to training models that are robust to outliers from the start, without sacrificing performance or efficiency. Specifically, OSP focuses on the following goals:
1. 📈**Scaling to production-level training requirements**<br/>
Prior methods for quantization-friendly pre-training are often limited to small-scale experiments (e.g., models under 1B parameters or 100B tokens). In contrast, we train a 1.4B-parameter model on 1 trillion tokens, demonstrating that OSP is effective at production scale.
2. ⚡**Maintaining computational efficiency comparable to standard training**<br/>
A method that prevents outliers but significantly reduces efficiency is unlikely to gain adoption. OSP introduces only a ~2% slowdown while reducing GPU memory usage, making it appealing for those seeking to train quantization-friendly foundation models from scratch.
3. 🧩**Ensuring full compatibility with existing inference pipelines**<br/>
We prioritize compatibility with widely adopted inference frameworks such as vLLM and SGLang. Rather than introducing architectural changes that break compatibility, OSP preserves computational invariance, allowing models to be directly integrated into existing pipelines without additional effort.
## Model Checkpoints
### Final Models
The models were trained on 1 trillion tokens, following the pre-training recipe of [SmolLM](https://huggingface.co/blog/smollm). Specifically, training was conducted using the [smollm-corpus](https://huggingface.co/datasets/HuggingFaceTB/smollm-corpus), a mixture of FineWeb-Edu, Cosmopedia, and Python-Edu.
- [🤗 OSP-1.4B-1T-Adam](https://huggingface.co/dmis-lab/OSP-1.4B-1T-Adam): Trained on the standard Adam optimizer, without any modifications.
- [🤗 OSP-1.4B-1T-Muon-SSNorm-EmbProj](https://huggingface.co/dmis-lab/OSP-1.4B-1T-Muon-SSNorm-EmbProj): Trained on the OSP framework. This is our final model.
### Ablation Models
<table>
<thead>
<tr>
<th rowspan="2">Model</th>
<th rowspan="2">Optimizer</th>
<th rowspan="2">SSNorm</th>
<th rowspan="2">EmbProj</th>
<th rowspan="2">Ex. Kurt.</th>
<th rowspan="2">Had.</th>
<!-- <th colspan="2">16-16-16</th> -->
<th colspan="2">4-4-4</th>
</tr>
<tr>
<!-- <th>Avg.</th>
<th>PPL</th> -->
<!-- <th>Avg.</th>
<th>PPL</th>
<th>Avg.</th>
<th>PPL</th>
<th>Avg.</th>
<th>PPL</th> -->
<th>Avg.</th>
<th>PPL</th>
</tr>
</thead>
<tbody>
<tr>
<td><a href="https://huggingface.co/dmis-lab/OSP-1.4B-100B-Adam">🤗 OSP-1.4B-100B-Adam</a></td>
<td>Adam</td>
<td>✗</td>
<td>✗</td>
<td>1818.56</td>
<td>✗<br>✔</td>
<!-- <td>41.5<br>41.5</td>
<td>11.4<br>11.4</td> -->
<!-- <td>39.7<br>40.2</td>
<td>21.6<br>22.3</td>
<td>39.7<br>40.3</td>
<td>21.6<br>22.3</td>
<td>26.5<br>27.2</td>
<td>1e5<br>3e4</td> -->
<td>26.8<br>26.9</td>
<td>8e4<br>3e4</td>
</tr>
<tr>
<td><a href="https://huggingface.co/dmis-lab/OSP-1.4B-100B-Muon-Only">🤗 OSP-1.4B-100B-Muon-Only</a></td>
<td>Muon†<br/>(w/o Adam)</td>
<td>✗</td>
<td>✗</td>
<td>361.35</td>
<td>✗<br>✔</td>
<!-- <td>41.0<br>41.0</td>
<td>11.7<br>11.7</td> -->
<!-- <td>38.4<br>37.5</td>
<td>14.8<br>15.4</td>
<td>38.3<br>37.5</td>
<td>14.8<br>15.4</td>
<td>26.3<br>33.3</td>
<td>1e6<br>24.5</td> -->
<td>26.3<br>33.1</td>
<td>8e5<br>24.8</td>
</tr>
<tr>
<td><a href="https://huggingface.co/dmis-lab/OSP-1.4B-100B-Muon">🤗 OSP-1.4B-100B-Muon</a></td>
<td>Muon</td>
<td>✗</td>
<td>✗</td>
<td>1575.12</td>
<td>✗<br>✔</td>
<!-- <td>41.5<br>41.5</td>
<td>11.4<br>11.4</td> -->
<!-- <td>40.0<br>40.6</td>
<td>13.8<br>12.9</td>
<td>40.0<br>40.6</td>
<td>13.8<br>12.9</td>
<td>29.4<br>38.6</td>
<td>934.3<br>15.7</td> -->
<td>29.0<br>38.4</td>
<td>1e4<br>15.8</td>
</tr>
<tr>
<td><a href="https://huggingface.co/dmis-lab/OSP-1.4B-100B-Muon-SSNorm">🤗 OSP-1.4B-100B-Muon-SSNorm</a></td>
<td>Muon</td>
<td>✔</td>
<td>✗</td>
<td>66.69</td>
<td>✗<br>✔</td>
<!-- <td><strong>41.8</strong><br><strong>41.8</strong></td>
<td><strong>11.2</strong><br><strong>11.2</strong></td> -->
<!-- <td><strong>41.0</strong><br><strong>40.8</strong></td>
<td>12.4<br>12.2</td>
<td><strong>40.9</strong><br><strong>40.8</strong></td>
<td>12.4<br>12.2</td>
<td>36.6<br>38.6</td>
<td>43.3<br>33.7</td> -->
<td>36.4<br>38.3</td>
<td>44.2<br>34.1</td>
</tr>
<tr>
<td><a href="https://huggingface.co/dmis-lab/OSP-1.4B-100B-Muon-EmbProj">🤗 OSP-1.4B-100B-Muon-EmbProj</a></td>
<td>Muon</td>
<td>✗</td>
<td>✔</td>
<td>703.23</td>
<td>✗<br>✔</td>
<!-- <td>40.0<br>40.0</td>
<td>12.3<br>12.3</td> -->
<!-- <td>38.4<br>39.2</td>
<td>14.8<br>13.9</td>
<td>38.4<br>39.3</td>
<td>14.8<br>13.9</td>
<td>31.0<br>36.3</td>
<td>99.7<br>22.1</td> -->
<td>30.4<br>36.2</td>
<td>114.6<br>22.3</td>
</tr>
<tr>
<td><a href="https://huggingface.co/dmis-lab/OSP-1.4B-100B-Muon-SSNorm-EmbProj">🤗 OSP-1.4B-100B-Muon-SSNorm-EmbProj</a></td>
<td>Muon</td>
<td>✔</td>
<td>✔</td>
<td><strong>0.04</strong></td>
<td>✗<br>✔</td>
<!-- <td>41.4<br>41.4</td>
<td><strong>11.2</strong><br><strong>11.2</strong></td> -->
<!-- <td>40.6<br>40.5</td>
<td><strong>12.2</strong><br><strong>12.1</strong></td>
<td>40.6<br>40.5</td>
<td><strong>12.2</strong><br><strong>12.1</strong></td>
<td><strong>37.9</strong><br><strong>39.1</strong></td>
<td><strong>19.4</strong><br><strong>13.4</strong></td> -->
<td><strong>37.5</strong><br><strong>38.9</strong></td>
<td><strong>19.6</strong><br><strong>13.5</strong></td>
</tr>
</tbody>
</table>
†Model configuration that disables decoupled embedding optimization by training with Muon optimizer without Adam optimization on embedding layers
## Training
### Model
- Architecture: Llama
- Pretraining tokens: 1 trillion tokens
- Precision: bfloat16
### Hardware
- TPUs: TPU-v4-512 Pod Slice (supported by [TRC Program](https://sites.research.google/trc/about/))
### Software
- Training Framework: [Jax](https://github.com/jax-ml/jax), [Flax](https://github.com/google/flax)
## Disclaimer
This model family was trained to demonstrate the effectiveness of eliminating outlier occurrences and improving quantization-friendliness. All models are base models, i.e., no instruction tuning or human alignment was applied. These models are not intended for chatting, conversation, or assistant purposes. They may contain toxic or harmful content. Their best use is for evaluating performance degradation on benchmarks after low-bit quantization.
## Citation
```bibtex
@article{park2025osp,
title={Outlier-Safe Pre-Training for Robust 4-Bit Quantization of Large Language Models},
author={Jungwoo Park and Taewhoo Lee and Chanwoong Yoon and Hyeon Hwang and Jaewoo Kang},
year={2025},
eprint={2506.19697},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2506.19697},
}
``` |