Advancing Model Refinement: Muon-Optimized Distillation and Quantization for LLM Deployment

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Sander, Jacob, Jalaian, Brian, Dasari, Venkat R.
Natura: Preprint
Pubblicazione: 2026
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866909990810288128
author Sander, Jacob
Jalaian, Brian
Dasari, Venkat R.
author_facet Sander, Jacob
Jalaian, Brian
Dasari, Venkat R.
contents Large Language Models (LLMs) enable advanced natural language processing but face deployment challenges on resource-constrained edge devices due to high computational, memory, and energy demands. Optimizing these models requires addressing three key challenges: acquiring task-specific data, fine-tuning for performance, and compressing models to accelerate inference while reducing resource demands. We propose an integrated framework combining GPTQ-based quantization, low-rank adaptation (LoRA), and a specialized data distillation process to significantly reduce model size and complexity while preserving or enhancing task-specific performance. By leveraging data distillation, knowledge distillation via Kullback-Leibler divergence, Bayesian hyperparameter optimization, and the Muon optimizer, our pipeline achieves up to 2x memory compression (e.g., reducing a 6GB model to 3GB) and enables efficient inference for specialized tasks. Empirical results demonstrate superior performance on standard LLM benchmarks compared to GPTQ quantization alone, with the Muon optimizer notably enhancing fine-tuned models' resistance to accuracy decay during quantization.
format Preprint
id arxiv_https___arxiv_org_abs_2601_09865
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Advancing Model Refinement: Muon-Optimized Distillation and Quantization for LLM Deployment
Sander, Jacob
Jalaian, Brian
Dasari, Venkat R.
Machine Learning
Artificial Intelligence
Large Language Models (LLMs) enable advanced natural language processing but face deployment challenges on resource-constrained edge devices due to high computational, memory, and energy demands. Optimizing these models requires addressing three key challenges: acquiring task-specific data, fine-tuning for performance, and compressing models to accelerate inference while reducing resource demands. We propose an integrated framework combining GPTQ-based quantization, low-rank adaptation (LoRA), and a specialized data distillation process to significantly reduce model size and complexity while preserving or enhancing task-specific performance. By leveraging data distillation, knowledge distillation via Kullback-Leibler divergence, Bayesian hyperparameter optimization, and the Muon optimizer, our pipeline achieves up to 2x memory compression (e.g., reducing a 6GB model to 3GB) and enables efficient inference for specialized tasks. Empirical results demonstrate superior performance on standard LLM benchmarks compared to GPTQ quantization alone, with the Muon optimizer notably enhancing fine-tuned models' resistance to accuracy decay during quantization.
title Advancing Model Refinement: Muon-Optimized Distillation and Quantization for LLM Deployment
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2601.09865