---
title: "LoRA and QLoRA fine-tuning | Skills We Assign For | Azendo"
description: "LoRA and QLoRA fine-tuning — adapting models affordably, and the Azendo roles assigned for it."
url: "https://azendo.co/skills/lora-and-qlora-fine-tuning/"
---

[Skills](https://azendo.co/skills/) AI and machine learning 

# LoRA and QLoRA fine-tuning.

LoRA fine-tunes a model by training small low-rank adapter matrices while the base weights stay frozen. QLoRA adds quantisation of the base model, allowing large models to be adapted on a single GPU.

## Where LoRA and QLoRA fine-tuning fits on a long engagement.

These techniques changed who can fine-tune. Full fine-tuning of a large model requires multiple high-memory GPUs; LoRA trains a small fraction of the parameters and QLoRA quantises the frozen base, bringing adaptation of a substantial model within reach of a single accessible GPU.

Adapters are small and swappable, which is architecturally useful. Several task-specific adapters can share one base model in memory, so serving five specialised variants does not mean five full model deployments.

## What an assigned team does with LoRA and QLoRA fine-tuning.

Fine-tuning is the wrong first move more often than not. Better prompting and better retrieval usually deliver more than adaptation, and they are faster to iterate on. Fine-tuning earns its place for consistent formatting, domain vocabulary and behaviour that prompting cannot reliably produce.

Establishing which of those a problem actually is, before committing to a training pipeline, is the judgement scoped under [managed ai services](https://azendo.co/services/ai-engineering/).

## What we use LoRA and QLoRA fine-tuning for.

* Adaptation on accessible hardware Fine-tuning a substantial model on a single GPU rather than a cluster.
* Several adapters on one base Task-specific variants sharing a model in memory rather than deployed separately.
* Fine-tuning only where prompting cannot reach Retrieval and prompting exhausted first, because they iterate faster.

## How LoRA and QLoRA fine-tuning capacity is assigned.

Fine-tuning capacity is assigned inside AI work, with prompting and retrieval evaluated first because they are cheaper to iterate.

## Roles we assign LoRA and QLoRA fine-tuning for

* [LLM Engineer AI engineering](https://azendo.co/services/ai-engineering/llm-engineer/)

## Service lines it sits in

* [AI engineering](https://azendo.co/services/ai-engineering/)

Capacity is agreed as a committed monthly capacity across a discipline, not per skill.

## Related in ai and machine learning

* [PyTorch — skill we assign for](https://azendo.co/skills/pytorch/)
* [TensorFlow — skill we assign for](https://azendo.co/skills/tensorflow/)
* [scikit-learn — skill we assign for](https://azendo.co/skills/scikit-learn/)
* [Pandas — skill we assign for](https://azendo.co/skills/pandas/)
* [NumPy — skill we assign for](https://azendo.co/skills/numpy/)
* [embeddings — skill we assign for](https://azendo.co/skills/embeddings/)
* [function calling — skill we assign for](https://azendo.co/skills/function-calling/)
* [vLLM — skill we assign for](https://azendo.co/skills/vllm/)

## Tell us what your roadmap needs LoRA and QLoRA fine-tuning for.

A service delivery manager replies with the disciplines we would assign, the monthly capacity and what the first month looks like.

[All skills we assign for](https://azendo.co/skills/)