@inproceedings{shamshoum-etal-2025-compact,
title = "{C}omp{A}ct: Compressed Activations for Memory-Efficient {LLM} Training",
author = "Shamshoum, Yara and
Hodos, Nitzan and
Sieradzki, Yuval and
Schuster, Assaf",
editor = "Chiruzzo, Luis and
Ritter, Alan and
Wang, Lu",
booktitle = "Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)",
month = apr,
year = "2025",
address = "Albuquerque, New Mexico",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2025.naacl-long.71/",
doi = "10.18653/v1/2025.naacl-long.71",
pages = "1511--1524",
ISBN = "979-8-89176-189-6",
abstract = "We introduce CompAct, a technique that reduces peak memory utilization on GPU by 25-30{\%} for pretraining and 50{\%} for fine-tuning of LLMs. Peak device memory is a major limiting factor in training LLMs, with various recent works aiming to reduce model memory. However most works don{'}t target the largest component of allocated memory during training: the model{'}s compute graph, which is stored for the backward pass. By storing low-rank, compressed activations to be used in the backward pass we greatly reduce the required memory, unlike previous methods which only reduce optimizer overheads or the number of trained parameters. Our compression uses random projection matrices, thus avoiding additional memory overheads. Comparisons with previous techniques for either pretraining or fine-tuning show that CompAct substantially improves existing compute-performance tradeoffs. We expect CompAct{'}s savings to scale even higher for larger models."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="shamshoum-etal-2025-compact">
<titleInfo>
<title>CompAct: Compressed Activations for Memory-Efficient LLM Training</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yara</namePart>
<namePart type="family">Shamshoum</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nitzan</namePart>
<namePart type="family">Hodos</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yuval</namePart>
<namePart type="family">Sieradzki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Assaf</namePart>
<namePart type="family">Schuster</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2025-04</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Luis</namePart>
<namePart type="family">Chiruzzo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alan</namePart>
<namePart type="family">Ritter</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lu</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Albuquerque, New Mexico</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-189-6</identifier>
</relatedItem>
<abstract>We introduce CompAct, a technique that reduces peak memory utilization on GPU by 25-30% for pretraining and 50% for fine-tuning of LLMs. Peak device memory is a major limiting factor in training LLMs, with various recent works aiming to reduce model memory. However most works don’t target the largest component of allocated memory during training: the model’s compute graph, which is stored for the backward pass. By storing low-rank, compressed activations to be used in the backward pass we greatly reduce the required memory, unlike previous methods which only reduce optimizer overheads or the number of trained parameters. Our compression uses random projection matrices, thus avoiding additional memory overheads. Comparisons with previous techniques for either pretraining or fine-tuning show that CompAct substantially improves existing compute-performance tradeoffs. We expect CompAct’s savings to scale even higher for larger models.</abstract>
<identifier type="citekey">shamshoum-etal-2025-compact</identifier>
<identifier type="doi">10.18653/v1/2025.naacl-long.71</identifier>
<location>
<url>https://aclanthology.org/2025.naacl-long.71/</url>
</location>
<part>
<date>2025-04</date>
<extent unit="page">
<start>1511</start>
<end>1524</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T CompAct: Compressed Activations for Memory-Efficient LLM Training
%A Shamshoum, Yara
%A Hodos, Nitzan
%A Sieradzki, Yuval
%A Schuster, Assaf
%Y Chiruzzo, Luis
%Y Ritter, Alan
%Y Wang, Lu
%S Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)
%D 2025
%8 April
%I Association for Computational Linguistics
%C Albuquerque, New Mexico
%@ 979-8-89176-189-6
%F shamshoum-etal-2025-compact
%X We introduce CompAct, a technique that reduces peak memory utilization on GPU by 25-30% for pretraining and 50% for fine-tuning of LLMs. Peak device memory is a major limiting factor in training LLMs, with various recent works aiming to reduce model memory. However most works don’t target the largest component of allocated memory during training: the model’s compute graph, which is stored for the backward pass. By storing low-rank, compressed activations to be used in the backward pass we greatly reduce the required memory, unlike previous methods which only reduce optimizer overheads or the number of trained parameters. Our compression uses random projection matrices, thus avoiding additional memory overheads. Comparisons with previous techniques for either pretraining or fine-tuning show that CompAct substantially improves existing compute-performance tradeoffs. We expect CompAct’s savings to scale even higher for larger models.
%R 10.18653/v1/2025.naacl-long.71
%U https://aclanthology.org/2025.naacl-long.71/
%U https://doi.org/10.18653/v1/2025.naacl-long.71
%P 1511-1524
Markdown (Informal)
[CompAct: Compressed Activations for Memory-Efficient LLM Training](https://aclanthology.org/2025.naacl-long.71/) (Shamshoum et al., NAACL 2025)
ACL
- Yara Shamshoum, Nitzan Hodos, Yuval Sieradzki, and Assaf Schuster. 2025. CompAct: Compressed Activations for Memory-Efficient LLM Training. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 1511–1524, Albuquerque, New Mexico. Association for Computational Linguistics.