Fine-tuning large language models in resource-constrained environments: Methods and trade-offs
- 1Independent Researcher
Abstract
The rapid scaling of large language models has created significant barriers to fine-tuning, with standard approaches requiring computational resources far exceeding typical hardware constraints. A 7-billion parameter model demands roughly 100 to 120GB of memory for full fine-tuning under mixed-precision training, effectively limiting model customization to well-funded organizations with access to expensive multi-GPU infrastructure. This survey reviews resource-efficient fine-tuning methods through four complementary approaches: parameter-efficient fine-tuning (PEFT) methods including Low-Rank Adaptation (LoRA) and its variants, adapter-based architectures, and prompt-tuning strategies; quantization techniques such as QLoRA, GPTQ, and AWQ that reduce memory footprints through lower-precision representations; memory optimization strategies encompassing gradient checkpointing, FlashAttention, and CPU/disk offloading; and distributed and federated learning approaches that enable collaborative training across heterogeneous devices while preserving data privacy. Each approach reduces memory substantially, and stacked configurations reach roughly an order-of-magnitude reduction relative to the approximately 112 GB required for mixed-precision full fine-tuning of a 7-billion-parameter model (Rajbhandari et al., 2020). Parameter-efficient methods match or approach full fine-tuning quality on standard classification and natural language understanding benchmarks (Hu et al., 2021), with task-dependent gaps on complex generation. These figures are drawn from across the surveyed studies rather than a single controlled evaluation. Based on empirical evidence from recent literature, we develop a decision framework for method selection under varying resource constraints and identify open challenges including automated hyperparameter selection, joint quantization-adaptation optimization, and scalable federated learning. This survey aims to democratize access to large language model customization by providing practitioners and researchers with actionable guidance for fine-tuning under realistic resource limitations.
Preprint. This manuscript is under review at PeerJ Computer Science and has not completed peer review. Companion repository: https://github.com/jerryadamsfranklin/llm-efficient-finetuning-survey
Related
- CXRG-SVLM: a parameter-efficient small vision-language model for multi-view chest X-ray report generation in resource-constrained environments
- Generative decision intelligence: leveraging large language models for autonomous multi-objective resource orchestration in dynamic cloud environments
- Lightweight real-time messaging system for resource-constrained environments
- Automated Priority Rule Design for the Resource-Constrained Project Scheduling Problem: A Large Language Model-Guided Population-Based Search
- Multi-objective incremental path planning with learning-guided sampling and heading-constrained expansion for autonomous vehicles in dynamic occupancy grid environments
BibTeX
@misc{franklin2026finetuning,
title = {{Fine-tuning large language models in resource-constrained environments: Methods and trade-offs}},
author = {Jerry Adams Franklin},
howpublished = {Zenodo},
year = {2026},
month = oct,
doi = {10.5281/zenodo.23504006},
url = {https://zenodo.org/records/23504006},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Fine-tuning large language models in resource-constrained environments: Methods and trade-offs
AU - Jerry Adams Franklin
PY - 2026
DA - 2026/10/20
PB - Zenodo
DO - 10.5281/zenodo.23504006
UR - https://zenodo.org/records/23504006
AB - The rapid scaling of large language models has created significant barriers to fine-tuning, with standard approaches requiring computational resources far exceeding typical hardware constraints. A 7-billion parameter model demands roughly 100 to 120GB of memory for full fine-tuning under mixed-precision training, effectively limiting model customization to well-funded organizations with access to expensive multi-GPU infrastructure. This survey reviews resource-efficient fine-tuning methods through four complementary approaches: parameter-efficient fine-tuning (PEFT) methods including Low-Rank Adaptation (LoRA) and its variants, adapter-based architectures, and prompt-tuning strategies; quantization techniques such as QLoRA, GPTQ, and AWQ that reduce memory footprints through lower-precision representations; memory optimization strategies encompassing gradient checkpointing, FlashAttention, and CPU/disk offloading; and distributed and federated learning approaches that enable collaborative training across heterogeneous devices while preserving data privacy. Each approach reduces memory substantially, and stacked configurations reach roughly an order-of-magnitude reduction relative to the approximately 112 GB required for mixed-precision full fine-tuning of a 7-billion-parameter model (Rajbhandari et al., 2020). Parameter-efficient methods match or approach full fine-tuning quality on standard classification and natural language understanding benchmarks (Hu et al., 2021), with task-dependent gaps on complex generation. These figures are drawn from across the surveyed studies rather than a single controlled evaluation. Based on empirical evidence from recent literature, we develop a decision framework for method selection under varying resource constraints and identify open challenges including automated hyperparameter selection, joint quantization-adaptation optimization, and scalable federated learning. This survey aims to democratize access to large language model customization by providing practitioners and researchers with actionable guidance for fine-tuning under realistic resource limitations. Preprint. This manuscript is under review at PeerJ Computer Science and has not completed peer review. Companion repository: https://github.com/jerryadamsfranklin/llm-efficient-finetuning-survey
ER -