@inproceedings{xu-etal-2026-mf,
title = "{MF}-{QAT}: Multi-Format Quantization-Aware Training for Elastic Inference",
author = "Xu, Zifei and
Sharify, Sayeh and
Mostafa, Hesham",
editor = "Zhang, Yue and
Xu, Ruifeng and
Purwarianti, Ayu and
Rong, Tong and
Suzuki, Jun",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {IJCNLP}-{AACL} 2026",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-aacl.110/",
pages = "2168--2188",
ISBN = "979-8-89176-444-6",
abstract = "Quantization-aware training (QAT) is typically performed for a single target numeric format, while practical deployments often need to choose numerical precision at inference time based on hardware support or runtime constraints. We study multi-format QAT, which fine-tunes a pretrained model for multiple precisions within a microscaling format family. Across the evaluated language models, it achieves accuracy close to single-format QAT at most target precisions and generalizes to intermediate precisions unseen during fine-tuning, although the lowest-precision MXINT setting remains challenging. To make these models deployable from a single quantized checkpoint, we propose Slice-and-Scale, which converts a higher-precision MXINT or MXFP anchor to lower precisions by transforming both element values and shared block scales, without access to the original full-precision weights. Fine-tuning through this conversion produces an anchor checkpoint whose converted variants closely track standard multi-format QAT in language-modeling evaluations. Together, multi-format QAT and Slice-and-Scale enable precision selection from one stored checkpoint within each format family."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="xu-etal-2026-mf">
<titleInfo>
<title>MF-QAT: Multi-Format Quantization-Aware Training for Elastic Inference</title>
</titleInfo>
<name type="personal">
<namePart type="given">Zifei</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sayeh</namePart>
<namePart type="family">Sharify</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hesham</namePart>
<namePart type="family">Mostafa</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruifeng</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ayu</namePart>
<namePart type="family">Purwarianti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tong</namePart>
<namePart type="family">Rong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jun</namePart>
<namePart type="family">Suzuki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Hengqin, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-444-6</identifier>
</relatedItem>
<abstract>Quantization-aware training (QAT) is typically performed for a single target numeric format, while practical deployments often need to choose numerical precision at inference time based on hardware support or runtime constraints. We study multi-format QAT, which fine-tunes a pretrained model for multiple precisions within a microscaling format family. Across the evaluated language models, it achieves accuracy close to single-format QAT at most target precisions and generalizes to intermediate precisions unseen during fine-tuning, although the lowest-precision MXINT setting remains challenging. To make these models deployable from a single quantized checkpoint, we propose Slice-and-Scale, which converts a higher-precision MXINT or MXFP anchor to lower precisions by transforming both element values and shared block scales, without access to the original full-precision weights. Fine-tuning through this conversion produces an anchor checkpoint whose converted variants closely track standard multi-format QAT in language-modeling evaluations. Together, multi-format QAT and Slice-and-Scale enable precision selection from one stored checkpoint within each format family.</abstract>
<identifier type="citekey">xu-etal-2026-mf</identifier>
<location>
<url>https://aclanthology.org/2026.findings-aacl.110/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>2168</start>
<end>2188</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T MF-QAT: Multi-Format Quantization-Aware Training for Elastic Inference
%A Xu, Zifei
%A Sharify, Sayeh
%A Mostafa, Hesham
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-444-6
%F xu-etal-2026-mf
%X Quantization-aware training (QAT) is typically performed for a single target numeric format, while practical deployments often need to choose numerical precision at inference time based on hardware support or runtime constraints. We study multi-format QAT, which fine-tunes a pretrained model for multiple precisions within a microscaling format family. Across the evaluated language models, it achieves accuracy close to single-format QAT at most target precisions and generalizes to intermediate precisions unseen during fine-tuning, although the lowest-precision MXINT setting remains challenging. To make these models deployable from a single quantized checkpoint, we propose Slice-and-Scale, which converts a higher-precision MXINT or MXFP anchor to lower precisions by transforming both element values and shared block scales, without access to the original full-precision weights. Fine-tuning through this conversion produces an anchor checkpoint whose converted variants closely track standard multi-format QAT in language-modeling evaluations. Together, multi-format QAT and Slice-and-Scale enable precision selection from one stored checkpoint within each format family.
%U https://aclanthology.org/2026.findings-aacl.110/
%P 2168-2188
Markdown (Informal)
[MF-QAT: Multi-Format Quantization-Aware Training for Elastic Inference](https://aclanthology.org/2026.findings-aacl.110/) (Xu et al., Findings 2026)
ACL