@inproceedings{bajwa-2026-comet,
title = "{COMET} Sensitivity-Guided Mixed Precision Quantization for the {WMT}26 Model Compression Task",
author = "Bajwa, Angad Ripudaman Singh",
editor = "Haddow, Barry and
Kocmi, Tom and
Koehn, Philipp and
Monz, Christof",
booktitle = "Proceedings of the Eleventh Conference on Machine Translation",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.wmt-main.133/",
pages = "1973--1978",
ISBN = "979-8-89176-474-3",
abstract = "The WMT26 Model Compression shared task asks participants to reduce the footprint of a general-purpose large language model for machine translation (MT) deployment while preserving translation quality, under a fixed evaluation budget of model size, VRAM usage, and inference speed. We report a systematic comparison of post-training compression strategies applied to the constrained track base model, google/gemma-3-12b-it, across the three required language directions (Czech{--}German, English{--}Chinese (Simplified), English{--}Arabic (Egyptian)). We evaluate an uncompressed BF16 baseline and a vLLM-served engine-level control against six compression strategies: two bitsandbytes baselines (8-bit and 4-bit), two calibrated weight-only quantization methods (AWQ and GPTQ, both W4A16), vLLM{'}s native dynamic FP8 weight quantization, and a novel COMET-sensitivity-guided mixed-precision scheme that assigns each of the 48 decoder layers an independent BF16/INT8/INT4 tier via a size-budget knapsack. Quality is measured with chrF and three COMET variants. Speed is full-process wall-clock time (including model load) on a held-out blind test set, and size is the on-disk safetensors footprint. Our mixed-precision submission achieves the best quality-size-speed trade-off in our pool: it is the fastest system measured (1.76 sentences/sec, batch 16), within 0.4 chrF and a few COMET points of the uncompressed baseline across all three language pairs, at roughly 40{\%} of the baseline{'}s on-disk size (9.15GB vs 23GB)."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="bajwa-2026-comet">
<titleInfo>
<title>COMET Sensitivity-Guided Mixed Precision Quantization for the WMT26 Model Compression Task</title>
</titleInfo>
<name type="personal">
<namePart type="given">Angad</namePart>
<namePart type="given">Ripudaman</namePart>
<namePart type="given">Singh</namePart>
<namePart type="family">Bajwa</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the Eleventh Conference on Machine Translation</title>
</titleInfo>
<name type="personal">
<namePart type="given">Barry</namePart>
<namePart type="family">Haddow</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tom</namePart>
<namePart type="family">Kocmi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Philipp</namePart>
<namePart type="family">Koehn</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Christof</namePart>
<namePart type="family">Monz</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-474-3</identifier>
</relatedItem>
<abstract>The WMT26 Model Compression shared task asks participants to reduce the footprint of a general-purpose large language model for machine translation (MT) deployment while preserving translation quality, under a fixed evaluation budget of model size, VRAM usage, and inference speed. We report a systematic comparison of post-training compression strategies applied to the constrained track base model, google/gemma-3-12b-it, across the three required language directions (Czech–German, English–Chinese (Simplified), English–Arabic (Egyptian)). We evaluate an uncompressed BF16 baseline and a vLLM-served engine-level control against six compression strategies: two bitsandbytes baselines (8-bit and 4-bit), two calibrated weight-only quantization methods (AWQ and GPTQ, both W4A16), vLLM’s native dynamic FP8 weight quantization, and a novel COMET-sensitivity-guided mixed-precision scheme that assigns each of the 48 decoder layers an independent BF16/INT8/INT4 tier via a size-budget knapsack. Quality is measured with chrF and three COMET variants. Speed is full-process wall-clock time (including model load) on a held-out blind test set, and size is the on-disk safetensors footprint. Our mixed-precision submission achieves the best quality-size-speed trade-off in our pool: it is the fastest system measured (1.76 sentences/sec, batch 16), within 0.4 chrF and a few COMET points of the uncompressed baseline across all three language pairs, at roughly 40% of the baseline’s on-disk size (9.15GB vs 23GB).</abstract>
<identifier type="citekey">bajwa-2026-comet</identifier>
<location>
<url>https://aclanthology.org/2026.wmt-main.133/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>1973</start>
<end>1978</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T COMET Sensitivity-Guided Mixed Precision Quantization for the WMT26 Model Compression Task
%A Bajwa, Angad Ripudaman Singh
%Y Haddow, Barry
%Y Kocmi, Tom
%Y Koehn, Philipp
%Y Monz, Christof
%S Proceedings of the Eleventh Conference on Machine Translation
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-474-3
%F bajwa-2026-comet
%X The WMT26 Model Compression shared task asks participants to reduce the footprint of a general-purpose large language model for machine translation (MT) deployment while preserving translation quality, under a fixed evaluation budget of model size, VRAM usage, and inference speed. We report a systematic comparison of post-training compression strategies applied to the constrained track base model, google/gemma-3-12b-it, across the three required language directions (Czech–German, English–Chinese (Simplified), English–Arabic (Egyptian)). We evaluate an uncompressed BF16 baseline and a vLLM-served engine-level control against six compression strategies: two bitsandbytes baselines (8-bit and 4-bit), two calibrated weight-only quantization methods (AWQ and GPTQ, both W4A16), vLLM’s native dynamic FP8 weight quantization, and a novel COMET-sensitivity-guided mixed-precision scheme that assigns each of the 48 decoder layers an independent BF16/INT8/INT4 tier via a size-budget knapsack. Quality is measured with chrF and three COMET variants. Speed is full-process wall-clock time (including model load) on a held-out blind test set, and size is the on-disk safetensors footprint. Our mixed-precision submission achieves the best quality-size-speed trade-off in our pool: it is the fastest system measured (1.76 sentences/sec, batch 16), within 0.4 chrF and a few COMET points of the uncompressed baseline across all three language pairs, at roughly 40% of the baseline’s on-disk size (9.15GB vs 23GB).
%U https://aclanthology.org/2026.wmt-main.133/
%P 1973-1978
Markdown (Informal)
[COMET Sensitivity-Guided Mixed Precision Quantization for the WMT26 Model Compression Task](https://aclanthology.org/2026.wmt-main.133/) (Bajwa, WMT 2026)
ACL