@inproceedings{kodavanti-etal-2026-unlocking,
title = "Unlocking the Edge deployment and ondevice acceleration of multi-{L}o{RA} enabled one-for-all foundational {LLM}",
author = "Kodavanti, Sravanth and
Vajrala, Sowmya and
Miriyala, Srinivas Soumitri and
Tiwari, Utsav and
Kumar, Uttam and
Mahawar, Utkarsh Kumar and
Singh, Achal Pratap and
D, Arya and
Mutyala, Narendra and
Rajendiran, Vikram Nelvoy and
Allur, Sharan Kumar and
Lee, Euntaik and
Kim, Dohyoung and
Lee, HyeonSu and
Cho, Gyusung and
Kim, JungBae",
editor = "Liakata, Maria and
Moreira, Viviane P. and
Zhang, Jiajun and
Jurgens, David",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {ACL} 2026",
month = jul,
year = "2026",
address = "San Diego, California, United States",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-acl.2106/",
doi = "10.18653/v1/2026.findings-acl.2106",
pages = "42443--42455",
ISBN = "979-8-89176-395-1",
abstract = "Deploying large language models (LLMs) on smartphones poses significant engineering challenges due to stringent constraints on memory, latency, and runtime flexibility. In this work, we present a hardware-aware framework for efficient on-device inference of a LLaMA-based multilingual foundation model supporting multiple use cases on Samsung Galaxy S24 and S25 devices with SM8650 and SM8750 Qualcomm chipsets respectively. Our approach integrates application-specific LoRAs as runtime inputs to a single frozen inference graph, enabling dynamic task switching without recompilation or memory overhead. We further introduce a multi-stream decoding mechanism that concurrently generates stylistic variations{---}such as formal, polite, or jovial responses{---}within a single forward pass, reducing latency by up to 6{\texttimes}. To accelerate token generation, we apply Dynamic Self-Speculative Decoding (DS2D), a tree-based strategy that predicts future tokens without requiring a draft model, yielding up to 2.3{\texttimes} speedup in decode time. Combined with quantization to INT4 and architecture-level optimizations, our system achieves 4{--}6{\texttimes} overall improvements in memory and latency while maintaining accuracy across 9 languages and 8 tasks. These results demonstrate practical feasibility of deploying multi-use-case LLMs on edge devices, advancing the commercial viability of Generative AI in mobile platforms."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="kodavanti-etal-2026-unlocking">
<titleInfo>
<title>Unlocking the Edge deployment and ondevice acceleration of multi-LoRA enabled one-for-all foundational LLM</title>
</titleInfo>
<name type="personal">
<namePart type="given">Sravanth</namePart>
<namePart type="family">Kodavanti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sowmya</namePart>
<namePart type="family">Vajrala</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Srinivas</namePart>
<namePart type="given">Soumitri</namePart>
<namePart type="family">Miriyala</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Utsav</namePart>
<namePart type="family">Tiwari</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Uttam</namePart>
<namePart type="family">Kumar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Utkarsh</namePart>
<namePart type="given">Kumar</namePart>
<namePart type="family">Mahawar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Achal</namePart>
<namePart type="given">Pratap</namePart>
<namePart type="family">Singh</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Arya</namePart>
<namePart type="family">D</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Narendra</namePart>
<namePart type="family">Mutyala</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Vikram</namePart>
<namePart type="given">Nelvoy</namePart>
<namePart type="family">Rajendiran</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sharan</namePart>
<namePart type="given">Kumar</namePart>
<namePart type="family">Allur</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Euntaik</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Dohyoung</namePart>
<namePart type="family">Kim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">HyeonSu</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Gyusung</namePart>
<namePart type="family">Cho</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">JungBae</namePart>
<namePart type="family">Kim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-07</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: ACL 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Maria</namePart>
<namePart type="family">Liakata</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Viviane</namePart>
<namePart type="given">P</namePart>
<namePart type="family">Moreira</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jiajun</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">David</namePart>
<namePart type="family">Jurgens</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">San Diego, California, United States</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-395-1</identifier>
</relatedItem>
<abstract>Deploying large language models (LLMs) on smartphones poses significant engineering challenges due to stringent constraints on memory, latency, and runtime flexibility. In this work, we present a hardware-aware framework for efficient on-device inference of a LLaMA-based multilingual foundation model supporting multiple use cases on Samsung Galaxy S24 and S25 devices with SM8650 and SM8750 Qualcomm chipsets respectively. Our approach integrates application-specific LoRAs as runtime inputs to a single frozen inference graph, enabling dynamic task switching without recompilation or memory overhead. We further introduce a multi-stream decoding mechanism that concurrently generates stylistic variations—such as formal, polite, or jovial responses—within a single forward pass, reducing latency by up to 6×. To accelerate token generation, we apply Dynamic Self-Speculative Decoding (DS2D), a tree-based strategy that predicts future tokens without requiring a draft model, yielding up to 2.3× speedup in decode time. Combined with quantization to INT4 and architecture-level optimizations, our system achieves 4–6× overall improvements in memory and latency while maintaining accuracy across 9 languages and 8 tasks. These results demonstrate practical feasibility of deploying multi-use-case LLMs on edge devices, advancing the commercial viability of Generative AI in mobile platforms.</abstract>
<identifier type="citekey">kodavanti-etal-2026-unlocking</identifier>
<identifier type="doi">10.18653/v1/2026.findings-acl.2106</identifier>
<location>
<url>https://aclanthology.org/2026.findings-acl.2106/</url>
</location>
<part>
<date>2026-07</date>
<extent unit="page">
<start>42443</start>
<end>42455</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Unlocking the Edge deployment and ondevice acceleration of multi-LoRA enabled one-for-all foundational LLM
%A Kodavanti, Sravanth
%A Vajrala, Sowmya
%A Miriyala, Srinivas Soumitri
%A Tiwari, Utsav
%A Kumar, Uttam
%A Mahawar, Utkarsh Kumar
%A Singh, Achal Pratap
%A D, Arya
%A Mutyala, Narendra
%A Rajendiran, Vikram Nelvoy
%A Allur, Sharan Kumar
%A Lee, Euntaik
%A Kim, Dohyoung
%A Lee, HyeonSu
%A Cho, Gyusung
%A Kim, JungBae
%Y Liakata, Maria
%Y Moreira, Viviane P.
%Y Zhang, Jiajun
%Y Jurgens, David
%S Findings of the Association for Computational Linguistics: ACL 2026
%D 2026
%8 July
%I Association for Computational Linguistics
%C San Diego, California, United States
%@ 979-8-89176-395-1
%F kodavanti-etal-2026-unlocking
%X Deploying large language models (LLMs) on smartphones poses significant engineering challenges due to stringent constraints on memory, latency, and runtime flexibility. In this work, we present a hardware-aware framework for efficient on-device inference of a LLaMA-based multilingual foundation model supporting multiple use cases on Samsung Galaxy S24 and S25 devices with SM8650 and SM8750 Qualcomm chipsets respectively. Our approach integrates application-specific LoRAs as runtime inputs to a single frozen inference graph, enabling dynamic task switching without recompilation or memory overhead. We further introduce a multi-stream decoding mechanism that concurrently generates stylistic variations—such as formal, polite, or jovial responses—within a single forward pass, reducing latency by up to 6×. To accelerate token generation, we apply Dynamic Self-Speculative Decoding (DS2D), a tree-based strategy that predicts future tokens without requiring a draft model, yielding up to 2.3× speedup in decode time. Combined with quantization to INT4 and architecture-level optimizations, our system achieves 4–6× overall improvements in memory and latency while maintaining accuracy across 9 languages and 8 tasks. These results demonstrate practical feasibility of deploying multi-use-case LLMs on edge devices, advancing the commercial viability of Generative AI in mobile platforms.
%R 10.18653/v1/2026.findings-acl.2106
%U https://aclanthology.org/2026.findings-acl.2106/
%U https://doi.org/10.18653/v1/2026.findings-acl.2106
%P 42443-42455
Markdown (Informal)
[Unlocking the Edge deployment and ondevice acceleration of multi-LoRA enabled one-for-all foundational LLM](https://aclanthology.org/2026.findings-acl.2106/) (Kodavanti et al., Findings 2026)
ACL
- Sravanth Kodavanti, Sowmya Vajrala, Srinivas Soumitri Miriyala, Utsav Tiwari, Uttam Kumar, Utkarsh Kumar Mahawar, Achal Pratap Singh, Arya D, Narendra Mutyala, Vikram Nelvoy Rajendiran, Sharan Kumar Allur, Euntaik Lee, Dohyoung Kim, HyeonSu Lee, Gyusung Cho, and JungBae Kim. 2026. Unlocking the Edge deployment and ondevice acceleration of multi-LoRA enabled one-for-all foundational LLM. In Findings of the Association for Computational Linguistics: ACL 2026, pages 42443–42455, San Diego, California, United States. Association for Computational Linguistics.