Lost in Quantization: Activation Outliers Explain Language-Specific FP8 Sensitivity in Llama-3

Guilherme Silva; Pedro Silva; Matheus Peixoto; Gladston Moreira; Eduardo Luz

Lost in Quantization: Activation Outliers Explain Language-Specific FP8 Sensitivity in Llama-3

Guilherme Silva, Pedro Silva, Matheus Peixoto, Gladston Moreira, Eduardo Luz

Correct Metadata for

Use this form to create a GitHub issue with structured data describing the correction. You will need a GitHub account. Once you create that issue, the correction will be reviewed by a staff member.

⚠️ Mobile Users: Submitting this form to create a new issue will only work with github.com, not the GitHub Mobile app.

Important: The Anthology treat PDFs as authoritative. Please use this form only to correct data that is out of line with the PDF. See our corrections guidelines if you need to change the PDF.

Title Adjust the title. Retain tags such as <fixed-case>.

Authors Adjust author names and order to match the PDF.

Abstract Correct abstract if needed. Retain XML formatting tags such as <tex-math>. You may use <b>...</b> for bold, <i>...</i> for italic, and <url>...</url> for URLs.

Verification against PDF Ensure that the new title/authors match the snapshot below. (If there is no snapshot or it is too small, consult the PDF.)

Authors concatenated from the text boxes above:

ALL author names match the snapshot above—including middle initials, hyphens, and accents.

Abstract

Quantization is key for efficient LLM inference, but its language-specific effects are understudied. We compare INT8 and FP8 (E4M3) quantization for Meta-Llama-3-8B on English and Brazilian Portuguese (PT-BR). INT8 with outlier handling preserves perplexity in both languages, while naive FP8 casting degrades English far more than PT-BR (+18% vs. +3.9%). Activation analysis shows rarer, larger English spikes (>35) that are more prone to saturation under unscaled E4M3, whereas PT-BR activations are more concentrated. Our FP8 results reflect a naive casting stress test (no calibration/scaling), not an optimized FP8 recipe.

Anthology ID:: 2026.propor-1.108
Volume:: Proceedings of the 17th International Conference on Computational Processing of Portuguese (PROPOR 2026) - Vol. 1
Month:: April
Year:: 2026
Address:: Salvador, Brazil
Editors:: Marlo Souza, Iria de-Dios-Flores, Diana Santos, Larissa Freitas, Jackson Wilke da Cruz Souza, Eugénio Ribeiro
Venue:: PROPOR
SIG:
Publisher:: Association for Computational Linguistics
Note:
Pages:: 1044–1048
Language:
URL:: https://aclanthology.org/2026.propor-1.108/
DOI:
Bibkey:
Cite (ACL):: Guilherme Silva, Pedro Silva, Matheus Peixoto, Gladston Moreira, and Eduardo Luz. 2026. Lost in Quantization: Activation Outliers Explain Language-Specific FP8 Sensitivity in Llama-3. In Proceedings of the 17th International Conference on Computational Processing of Portuguese (PROPOR 2026) - Vol. 1, pages 1044–1048, Salvador, Brazil. Association for Computational Linguistics.
Cite (Informal):: Lost in Quantization: Activation Outliers Explain Language-Specific FP8 Sensitivity in Llama-3 (Silva et al., PROPOR 2026)
Copy Citation:
PDF:: https://aclanthology.org/2026.propor-1.108.pdf

PDF Cite Search Fix data

Export citation

BibTeX
MODS XML
Endnote
Preformatted

@inproceedings{silva-etal-2026-lost,
    title = "Lost in Quantization: Activation Outliers Explain Language-Specific {FP}8 Sensitivity in Llama-3",
    author = "Silva, Guilherme  and
      Silva, Pedro  and
      Peixoto, Matheus  and
      Moreira, Gladston  and
      Luz, Eduardo",
    editor = "Souza, Marlo  and
      de-Dios-Flores, Iria  and
      Santos, Diana  and
      Freitas, Larissa  and
      Souza, Jackson Wilke da Cruz  and
      Ribeiro, Eug{\'e}nio",
    booktitle = "Proceedings of the 17th International Conference on Computational Processing of {P}ortuguese ({PROPOR} 2026) - Vol. 1",
    month = apr,
    year = "2026",
    address = "Salvador, Brazil",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.propor-1.108/",
    pages = "1044--1048",
    ISBN = "979-8-89176-387-6",
    abstract = "Quantization is key for efficient LLM inference, but its language-specific effects are understudied. We compare INT8 and FP8 (E4M3) quantization for Meta-Llama-3-8B on English and Brazilian Portuguese (PT-BR). INT8 with outlier handling preserves perplexity in both languages, while naive FP8 casting degrades English far more than PT-BR (+18{\%} vs. +3.9{\%}). Activation analysis shows rarer, larger English spikes ($>35$) that are more prone to saturation under unscaled E4M3, whereas PT-BR activations are more concentrated. Our FP8 results reflect a naive casting stress test (no calibration/scaling), not an optimized FP8 recipe."
}

Download as File

<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="silva-etal-2026-lost">
    <titleInfo>
        <title>Lost in Quantization: Activation Outliers Explain Language-Specific FP8 Sensitivity in Llama-3</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Guilherme</namePart>
        <namePart type="family">Silva</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Pedro</namePart>
        <namePart type="family">Silva</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Matheus</namePart>
        <namePart type="family">Peixoto</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Gladston</namePart>
        <namePart type="family">Moreira</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Eduardo</namePart>
        <namePart type="family">Luz</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <originInfo>
        <dateIssued>2026-04</dateIssued>
    </originInfo>
    <typeOfResource>text</typeOfResource>
    <relatedItem type="host">
        <titleInfo>
            <title>Proceedings of the 17th International Conference on Computational Processing of Portuguese (PROPOR 2026) - Vol. 1</title>
        </titleInfo>
        <name type="personal">
            <namePart type="given">Marlo</namePart>
            <namePart type="family">Souza</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Iria</namePart>
            <namePart type="family">de-Dios-Flores</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Diana</namePart>
            <namePart type="family">Santos</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Larissa</namePart>
            <namePart type="family">Freitas</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Jackson</namePart>
            <namePart type="given">Wilke</namePart>
            <namePart type="given">da</namePart>
            <namePart type="given">Cruz</namePart>
            <namePart type="family">Souza</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Eugénio</namePart>
            <namePart type="family">Ribeiro</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <originInfo>
            <publisher>Association for Computational Linguistics</publisher>
            <place>
                <placeTerm type="text">Salvador, Brazil</placeTerm>
            </place>
        </originInfo>
        <genre authority="marcgt">conference publication</genre>
        <identifier type="isbn">979-8-89176-387-6</identifier>
    </relatedItem>
    <abstract>Quantization is key for efficient LLM inference, but its language-specific effects are understudied. We compare INT8 and FP8 (E4M3) quantization for Meta-Llama-3-8B on English and Brazilian Portuguese (PT-BR). INT8 with outlier handling preserves perplexity in both languages, while naive FP8 casting degrades English far more than PT-BR (+18% vs. +3.9%). Activation analysis shows rarer, larger English spikes (&gt;35) that are more prone to saturation under unscaled E4M3, whereas PT-BR activations are more concentrated. Our FP8 results reflect a naive casting stress test (no calibration/scaling), not an optimized FP8 recipe.</abstract>
    <identifier type="citekey">silva-etal-2026-lost</identifier>
    <location>
        <url>https://aclanthology.org/2026.propor-1.108/</url>
    </location>
    <part>
        <date>2026-04</date>
        <extent unit="page">
            <start>1044</start>
            <end>1048</end>
        </extent>
    </part>
</mods>
</modsCollection>

Download as File

%0 Conference Proceedings
%T Lost in Quantization: Activation Outliers Explain Language-Specific FP8 Sensitivity in Llama-3
%A Silva, Guilherme
%A Silva, Pedro
%A Peixoto, Matheus
%A Moreira, Gladston
%A Luz, Eduardo
%Y Souza, Marlo
%Y de-Dios-Flores, Iria
%Y Santos, Diana
%Y Freitas, Larissa
%Y Souza, Jackson Wilke da Cruz
%Y Ribeiro, Eugénio
%S Proceedings of the 17th International Conference on Computational Processing of Portuguese (PROPOR 2026) - Vol. 1
%D 2026
%8 April
%I Association for Computational Linguistics
%C Salvador, Brazil
%@ 979-8-89176-387-6
%F silva-etal-2026-lost
%X Quantization is key for efficient LLM inference, but its language-specific effects are understudied. We compare INT8 and FP8 (E4M3) quantization for Meta-Llama-3-8B on English and Brazilian Portuguese (PT-BR). INT8 with outlier handling preserves perplexity in both languages, while naive FP8 casting degrades English far more than PT-BR (+18% vs. +3.9%). Activation analysis shows rarer, larger English spikes (>35) that are more prone to saturation under unscaled E4M3, whereas PT-BR activations are more concentrated. Our FP8 results reflect a naive casting stress test (no calibration/scaling), not an optimized FP8 recipe.
%U https://aclanthology.org/2026.propor-1.108/
%P 1044-1048

Download as File

Markdown (Informal)

[Lost in Quantization: Activation Outliers Explain Language-Specific FP8 Sensitivity in Llama-3](https://aclanthology.org/2026.propor-1.108/) (Silva et al., PROPOR 2026)

Lost in Quantization: Activation Outliers Explain Language-Specific FP8 Sensitivity in Llama-3 (Silva et al., PROPOR 2026)

ACL

Guilherme Silva, Pedro Silva, Matheus Peixoto, Gladston Moreira, and Eduardo Luz. 2026. Lost in Quantization: Activation Outliers Explain Language-Specific FP8 Sensitivity in Llama-3. In Proceedings of the 17th International Conference on Computational Processing of Portuguese (PROPOR 2026) - Vol. 1, pages 1044–1048, Salvador, Brazil. Association for Computational Linguistics.