@inproceedings{wan-etal-2024-look,
title = "{LOOK}-{M}: Look-Once Optimization in {KV} Cache for Efficient Multimodal Long-Context Inference",
author = "Wan, Zhongwei and
Wu, Ziang and
Liu, Che and
Huang, Jinfa and
Zhu, Zhihong and
Jin, Peng and
Wang, Longyue and
Yuan, Li",
editor = "Al-Onaizan, Yaser and
Bansal, Mohit and
Chen, Yun-Nung",
booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2024",
month = nov,
year = "2024",
address = "Miami, Florida, USA",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2024.findings-emnlp.235/",
doi = "10.18653/v1/2024.findings-emnlp.235",
pages = "4065--4078",
abstract = "Long-context Multimodal Large Language Models (MLLMs) demand substantial computational resources for inference as the growth of their multimodal Key-Value (KV) cache, in response to increasing input lengths, challenges memory and time efficiency. Unlike single-modality LLMs that manage only textual contexts, the KV cache of long-context MLLMs includes representations from multiple images with temporal and spatial relationships and related textual contexts. The predominance of image tokens means traditional optimizations for LLMs' KV caches are unsuitable for multimodal long-context settings, and no prior works have addressed this challenge.In this work, we introduce **LOOK-M**, a pioneering, fine-tuning-free approach that efficiently reduces the multimodal KV cache size while maintaining performance comparable to a full cache. We observe that during prompt prefill, the model prioritizes more textual attention over image features, and based on the multimodal interaction observation, a new proposed text-prior method is explored to compress the KV cache. Furthermore, to mitigate the degradation of image contextual information, we propose several compensatory strategies using KV pairs merging. **LOOK-M** demonstrates that with a significant reduction in KV Cache memory usage, such as reducing it by **80{\%}** in some cases, it not only achieves approximately **1.3x** faster decoding but also maintains or even **enhances** performance across a variety of long context multimodal tasks."
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="wan-etal-2024-look">
<titleInfo>
<title>LOOK-M: Look-Once Optimization in KV Cache for Efficient Multimodal Long-Context Inference</title>
</titleInfo>
<name type="personal">
<namePart type="given">Zhongwei</namePart>
<namePart type="family">Wan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ziang</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Che</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jinfa</namePart>
<namePart type="family">Huang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhihong</namePart>
<namePart type="family">Zhu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Peng</namePart>
<namePart type="family">Jin</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Longyue</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Li</namePart>
<namePart type="family">Yuan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2024-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2024</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yaser</namePart>
<namePart type="family">Al-Onaizan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mohit</namePart>
<namePart type="family">Bansal</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yun-Nung</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Miami, Florida, USA</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Long-context Multimodal Large Language Models (MLLMs) demand substantial computational resources for inference as the growth of their multimodal Key-Value (KV) cache, in response to increasing input lengths, challenges memory and time efficiency. Unlike single-modality LLMs that manage only textual contexts, the KV cache of long-context MLLMs includes representations from multiple images with temporal and spatial relationships and related textual contexts. The predominance of image tokens means traditional optimizations for LLMs’ KV caches are unsuitable for multimodal long-context settings, and no prior works have addressed this challenge.In this work, we introduce **LOOK-M**, a pioneering, fine-tuning-free approach that efficiently reduces the multimodal KV cache size while maintaining performance comparable to a full cache. We observe that during prompt prefill, the model prioritizes more textual attention over image features, and based on the multimodal interaction observation, a new proposed text-prior method is explored to compress the KV cache. Furthermore, to mitigate the degradation of image contextual information, we propose several compensatory strategies using KV pairs merging. **LOOK-M** demonstrates that with a significant reduction in KV Cache memory usage, such as reducing it by **80%** in some cases, it not only achieves approximately **1.3x** faster decoding but also maintains or even **enhances** performance across a variety of long context multimodal tasks.</abstract>
<identifier type="citekey">wan-etal-2024-look</identifier>
<identifier type="doi">10.18653/v1/2024.findings-emnlp.235</identifier>
<location>
<url>https://aclanthology.org/2024.findings-emnlp.235/</url>
</location>
<part>
<date>2024-11</date>
<extent unit="page">
<start>4065</start>
<end>4078</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T LOOK-M: Look-Once Optimization in KV Cache for Efficient Multimodal Long-Context Inference
%A Wan, Zhongwei
%A Wu, Ziang
%A Liu, Che
%A Huang, Jinfa
%A Zhu, Zhihong
%A Jin, Peng
%A Wang, Longyue
%A Yuan, Li
%Y Al-Onaizan, Yaser
%Y Bansal, Mohit
%Y Chen, Yun-Nung
%S Findings of the Association for Computational Linguistics: EMNLP 2024
%D 2024
%8 November
%I Association for Computational Linguistics
%C Miami, Florida, USA
%F wan-etal-2024-look
%X Long-context Multimodal Large Language Models (MLLMs) demand substantial computational resources for inference as the growth of their multimodal Key-Value (KV) cache, in response to increasing input lengths, challenges memory and time efficiency. Unlike single-modality LLMs that manage only textual contexts, the KV cache of long-context MLLMs includes representations from multiple images with temporal and spatial relationships and related textual contexts. The predominance of image tokens means traditional optimizations for LLMs’ KV caches are unsuitable for multimodal long-context settings, and no prior works have addressed this challenge.In this work, we introduce **LOOK-M**, a pioneering, fine-tuning-free approach that efficiently reduces the multimodal KV cache size while maintaining performance comparable to a full cache. We observe that during prompt prefill, the model prioritizes more textual attention over image features, and based on the multimodal interaction observation, a new proposed text-prior method is explored to compress the KV cache. Furthermore, to mitigate the degradation of image contextual information, we propose several compensatory strategies using KV pairs merging. **LOOK-M** demonstrates that with a significant reduction in KV Cache memory usage, such as reducing it by **80%** in some cases, it not only achieves approximately **1.3x** faster decoding but also maintains or even **enhances** performance across a variety of long context multimodal tasks.
%R 10.18653/v1/2024.findings-emnlp.235
%U https://aclanthology.org/2024.findings-emnlp.235/
%U https://doi.org/10.18653/v1/2024.findings-emnlp.235
%P 4065-4078
Markdown (Informal)
[LOOK-M: Look-Once Optimization in KV Cache for Efficient Multimodal Long-Context Inference](https://aclanthology.org/2024.findings-emnlp.235/) (Wan et al., Findings 2024)
ACL
- Zhongwei Wan, Ziang Wu, Che Liu, Jinfa Huang, Zhihong Zhu, Peng Jin, Longyue Wang, and Li Yuan. 2024. LOOK-M: Look-Once Optimization in KV Cache for Efficient Multimodal Long-Context Inference. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 4065–4078, Miami, Florida, USA. Association for Computational Linguistics.