An empirical study of LLaMA3 quantization: from LLMs to MLLMs
Abstract The LLaMA family, a collection of foundation language models ranging from 7B to 65B parameters, has become one of the most powerful open-source large language models (LLMs) and the popular LLM backbone of multi-modal large language models (MLLMs), widely used in computer vision and natural...
Saved in:
| Main Authors: | , , , , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Springer
2024-12-01
|
| Series: | Visual Intelligence |
| Subjects: | |
| Online Access: | https://doi.org/10.1007/s44267-024-00070-x |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1841558993408884736 |
|---|---|
| author | Wei Huang Xingyu Zheng Xudong Ma Haotong Qin Chengtao Lv Hong Chen Jie Luo Xiaojuan Qi Xianglong Liu Michele Magno |
| author_facet | Wei Huang Xingyu Zheng Xudong Ma Haotong Qin Chengtao Lv Hong Chen Jie Luo Xiaojuan Qi Xianglong Liu Michele Magno |
| author_sort | Wei Huang |
| collection | DOAJ |
| description | Abstract The LLaMA family, a collection of foundation language models ranging from 7B to 65B parameters, has become one of the most powerful open-source large language models (LLMs) and the popular LLM backbone of multi-modal large language models (MLLMs), widely used in computer vision and natural language understanding tasks. In particular, LLaMA3 models have recently been released and have achieved impressive performance in various domains with super-large scale pre-training on over 15T tokens of data. Given the wide application of low-bit quantization for LLMs in resource-constrained scenarios, we explore LLaMA3’s capabilities when quantized to low bit-width. This exploration can potentially provide new insights and challenges for the low-bit quantization of LLaMA3 and other future LLMs, especially in addressing performance degradation issues that suffer in LLM compression. Specifically, we comprehensively evaluate the 10 existing post-training quantization and LoRA fine-tuning (LoRA-FT) methods of LLaMA3 on 1-8 bits and various datasets to reveal the low-bit quantization performance of LLaMA3. To uncover the capabilities of low-bit quantized MLLM, we assessed the performance of the LLaMA3-based LLaVA-Next-8B model under 2-4 ultra-low bits with post-training quantization methods. Our experimental results indicate that LLaMA3 still suffers from non-negligible degradation in linguistic and visual contexts, particularly under ultra-low bit widths. This highlights the significant performance gap at low bit-width that needs to be addressed in future developments. We expect that this empirical study will prove valuable in advancing future models, driving LLMs and MLLMs to achieve higher accuracy at lower bit to enhance practicality. |
| format | Article |
| id | doaj-art-02c0cca950694326b2c63289293d26aa |
| institution | Kabale University |
| issn | 2731-9008 |
| language | English |
| publishDate | 2024-12-01 |
| publisher | Springer |
| record_format | Article |
| series | Visual Intelligence |
| spelling | doaj-art-02c0cca950694326b2c63289293d26aa2025-01-05T12:50:15ZengSpringerVisual Intelligence2731-90082024-12-012111310.1007/s44267-024-00070-xAn empirical study of LLaMA3 quantization: from LLMs to MLLMsWei Huang0Xingyu Zheng1Xudong Ma2Haotong Qin3Chengtao Lv4Hong Chen5Jie Luo6Xiaojuan Qi7Xianglong Liu8Michele Magno9Department of Electrical and Electronic Engineering, The University of Hong KongSchool of Computer Science and Engineering, Beihang UniversitySchool of Computer Science and Engineering, Beihang UniversityDepartment of Information Technology and Electrical Engineering, ETH ZurichSchool of Computer Science and Engineering, Beihang UniversitySchool of Computer Science and Engineering, Beihang UniversitySchool of Computer Science and Engineering, Beihang UniversityDepartment of Electrical and Electronic Engineering, The University of Hong KongSchool of Computer Science and Engineering, Beihang UniversityDepartment of Information Technology and Electrical Engineering, ETH ZurichAbstract The LLaMA family, a collection of foundation language models ranging from 7B to 65B parameters, has become one of the most powerful open-source large language models (LLMs) and the popular LLM backbone of multi-modal large language models (MLLMs), widely used in computer vision and natural language understanding tasks. In particular, LLaMA3 models have recently been released and have achieved impressive performance in various domains with super-large scale pre-training on over 15T tokens of data. Given the wide application of low-bit quantization for LLMs in resource-constrained scenarios, we explore LLaMA3’s capabilities when quantized to low bit-width. This exploration can potentially provide new insights and challenges for the low-bit quantization of LLaMA3 and other future LLMs, especially in addressing performance degradation issues that suffer in LLM compression. Specifically, we comprehensively evaluate the 10 existing post-training quantization and LoRA fine-tuning (LoRA-FT) methods of LLaMA3 on 1-8 bits and various datasets to reveal the low-bit quantization performance of LLaMA3. To uncover the capabilities of low-bit quantized MLLM, we assessed the performance of the LLaMA3-based LLaVA-Next-8B model under 2-4 ultra-low bits with post-training quantization methods. Our experimental results indicate that LLaMA3 still suffers from non-negligible degradation in linguistic and visual contexts, particularly under ultra-low bit widths. This highlights the significant performance gap at low bit-width that needs to be addressed in future developments. We expect that this empirical study will prove valuable in advancing future models, driving LLMs and MLLMs to achieve higher accuracy at lower bit to enhance practicality.https://doi.org/10.1007/s44267-024-00070-xModel quantizationLarge language modelMulti-modalDeep learning |
| spellingShingle | Wei Huang Xingyu Zheng Xudong Ma Haotong Qin Chengtao Lv Hong Chen Jie Luo Xiaojuan Qi Xianglong Liu Michele Magno An empirical study of LLaMA3 quantization: from LLMs to MLLMs Visual Intelligence Model quantization Large language model Multi-modal Deep learning |
| title | An empirical study of LLaMA3 quantization: from LLMs to MLLMs |
| title_full | An empirical study of LLaMA3 quantization: from LLMs to MLLMs |
| title_fullStr | An empirical study of LLaMA3 quantization: from LLMs to MLLMs |
| title_full_unstemmed | An empirical study of LLaMA3 quantization: from LLMs to MLLMs |
| title_short | An empirical study of LLaMA3 quantization: from LLMs to MLLMs |
| title_sort | empirical study of llama3 quantization from llms to mllms |
| topic | Model quantization Large language model Multi-modal Deep learning |
| url | https://doi.org/10.1007/s44267-024-00070-x |
| work_keys_str_mv | AT weihuang anempiricalstudyofllama3quantizationfromllmstomllms AT xingyuzheng anempiricalstudyofllama3quantizationfromllmstomllms AT xudongma anempiricalstudyofllama3quantizationfromllmstomllms AT haotongqin anempiricalstudyofllama3quantizationfromllmstomllms AT chengtaolv anempiricalstudyofllama3quantizationfromllmstomllms AT hongchen anempiricalstudyofllama3quantizationfromllmstomllms AT jieluo anempiricalstudyofllama3quantizationfromllmstomllms AT xiaojuanqi anempiricalstudyofllama3quantizationfromllmstomllms AT xianglongliu anempiricalstudyofllama3quantizationfromllmstomllms AT michelemagno anempiricalstudyofllama3quantizationfromllmstomllms AT weihuang empiricalstudyofllama3quantizationfromllmstomllms AT xingyuzheng empiricalstudyofllama3quantizationfromllmstomllms AT xudongma empiricalstudyofllama3quantizationfromllmstomllms AT haotongqin empiricalstudyofllama3quantizationfromllmstomllms AT chengtaolv empiricalstudyofllama3quantizationfromllmstomllms AT hongchen empiricalstudyofllama3quantizationfromllmstomllms AT jieluo empiricalstudyofllama3quantizationfromllmstomllms AT xiaojuanqi empiricalstudyofllama3quantizationfromllmstomllms AT xianglongliu empiricalstudyofllama3quantizationfromllmstomllms AT michelemagno empiricalstudyofllama3quantizationfromllmstomllms |