LoRA_pipeline

The goal of this project is to provide a lightweight, modular system for fine-tuning and serving LLMs using techniques such as LoRA and QLoRA, enabling efficient training and deployment without full model retraining.

Scope

Implementation of LoRA-based fine-tuning workflows using Hugging Face PEFT Support for quantized training approaches (QLoRA) to reduce memory footprint Exploration of high-throughput and single-GPU fine-tuning strategies Integration considerations for serving multiple adapters (e.g. via vLLM) Early-stage research into scalable adapter usage and deployment patterns Purpose

This project is part of an ongoing effort to:

Reduce compute and memory requirements for LLM fine-tuning Enable rapid experimentation with domain-specific adaptations Bridge research (LoRA, QLoRA, S-LoRA, etc.) with practical production pipelines Status

Early-stage / research-driven.

The repository currently serves as a foundation for experimentation, benchmarking, and iterative development of efficient LLM fine-tuning systems.

Research

🤗 PEFT: State-of-the-art Parameter-Efficient Fine-Tuning.

https://github.com/huggingface/peft https://huggingface.co/docs/peft/en/package_reference/lora

Fine-tuning LLMs with PEFT and LoRA

https://www.youtube.com/watch?v=Us5ZFp16PaU

LORA: LOW-RANK ADAPTATION OF LARGE LANGUAGE MODELS 6 Oct 2021

https://arxiv.org/pdf/2106.09685

QLORA: Efficient Finetuning of Quantized LLMs 23 May 2023

https://arxiv.org/pdf/2305.14314

ASPEN: High-Throughput LoRA Fine-Tuning of Large Language Models with a Single GPU 5 Dec 2023

https://arxiv.org/pdf/2312.02515

LoRA+: Efficient Low Rank Adaptation of Large Models 19 Feb 2024

https://arxiv.org/abs/2402.12354

S-LORA: SERVING THOUSANDS OF CONCURRENT LORA ADAPTERS 7 Nov 2023

https://arxiv.org/pdf/2311.03285

Using LoRA adapters

This document shows you how to use LoRA adapters with vLLM https://docs.vllm.ai/en/latest/models/lora.html

wasm

https://github.com/ktock/container2wasm