A few years ago, fine-tuning a large language model meant booking time on a cluster of data-centre GPUs that cost more than a house. Today, thanks to two techniques called LoRA and QLoRA, you can adapt a capable 7-billion-parameter model on a single graphics card that fits inside a desktop tower in Lagos or Abuja. This is not a marketing exaggeration — it is a genuine shift in what the maths demands, and it has put serious model training within reach of Nigerian businesses, researchers, and developers.
This deep-dive explains exactly how that VRAM saving works, because understanding the mechanics is the only way to buy the right hardware without overspending. If you are still deciding whether to train locally at all, start with our guide on fine-tuning small LLMs locally, and if the very idea of "4-bit" sounds unfamiliar, our explainer on quantisation for local AI models sets the foundation for everything below.
Why full fine-tuning is so VRAM-hungry
To understand the breakthrough, you first need to understand the problem. When you train a model the traditional way — full fine-tuning — your GPU has to hold several things in memory at once, not just the model itself.
- The weights — the model's parameters. For a 7B model in 16-bit precision, that alone is roughly 14 GB.
- The gradients — for every weight you update, you compute and store a gradient. That roughly doubles the memory.
- The optimiser states — modern optimisers like Adam keep extra running statistics for each weight, often two more copies of the model's size.
Add it up and full fine-tuning typically needs somewhere around four times the model size in VRAM, before you even account for the data passing through. That is why full fine-tuning a 7B model has historically demanded data-centre cards. The weights are the smallest part of the bill — it is the gradients and optimiser states that break the budget.
How LoRA changes the equation
LoRA, which stands for Low-Rank Adaptation, is a parameter-efficient fine-tuning method — part of a family often abbreviated as PEFT. The idea is elegant. Instead of updating all of the model's weights, LoRA freezes the entire base model and inserts a small number of new, trainable "adapter" matrices alongside it. You train only those adapters.
Because the adapters are low-rank — deliberately small — you are training a tiny fraction of the total parameters, often well under one per cent. And here is the crucial part: gradients and optimiser states are only needed for the parameters you actually train. By freezing the base model, you eliminate almost the entire gradient and optimiser bill that made full fine-tuning so expensive. You still have to hold the frozen weights in memory, but you no longer pay the three-to-four-times multiplier on top of them.
How QLoRA goes one step further
LoRA solves the gradient and optimiser problem but leaves the frozen base weights sitting in memory at full size. QLoRA tackles that remaining cost head-on. It loads the frozen base model in 4-bit quantised precision — compressing the weights to a quarter of their 16-bit footprint — while training the LoRA adapters on top of it in higher precision.
This combination is the real breakthrough. By quantising the base weights and training only small adapters, QLoRA lets you fine-tune a model on a single consumer GPU that could never have held it for full fine-tuning. The headline figure quoted across the field is that QLoRA can comfortably fine-tune a 7B model on a 24 GB consumer card — the kind of hardware you can actually buy and run on NEPA power at home. For more on matching card capacity to model size, see what GPU VRAM you need in 2026.
Full fine-tuning vs LoRA vs QLoRA
Here is the practical comparison that matters when you are spending money:
- Full fine-tuning — trains every weight; needs weights plus gradients plus optimiser states (often four times the model size in VRAM); realistically needs multiple data-centre GPUs for a 7B model.
- LoRA — freezes the base model and trains small adapters; cuts gradient and optimiser cost to almost nothing, but still holds the full-size base weights; a 7B model becomes feasible on a high-end consumer card.
- QLoRA — freezes and quantises the base model to 4-bit, then trains adapters on top; lowest VRAM of all; fits a 7B model on a single 24 GB card with room to spare.
For a deeper look at how these choices scale with your model size and workload, our guide on GPU VRAM for training models in a business setting walks through the trade-offs in detail.
Understanding LoRA rank
The one tuning knob you will hear about most is rank. Rank controls how large the adapter matrices are. A higher rank gives the adapters slightly more capacity to learn — and costs a little more VRAM and compute — while a lower rank is leaner. The reassuring news is that most practical adaptation tasks work very well at modest ranks; you rarely need to push it high. Start small, measure the quality of your results, and only raise the rank if a specific task genuinely needs more capacity.
One lovely side effect of the LoRA approach: the trained adapters are small files, often just a few megabytes or tens of megabytes. You can keep one base model on disk and swap different adapters on top of it for different tasks. That makes them cheap to store, easy to share, and quick to back up — a real advantage when bandwidth and storage both cost money in Nigeria.
What fits on what, and rough Naira tiers
Translating the theory into a shopping list, here is roughly what each hardware tier lets you do with QLoRA. The used-GPU market is where most Nigerian builders find value — see our breakdown of the RTX 30 vs 40 used-GPU market before you buy.
- 16 GB VRAM (entry tier) — comfortably handles QLoRA on smaller models and a 7B model with careful settings; the most affordable serious starting point, typically a few hundred thousand Naira for a capable used card.
- 24 GB VRAM (the sweet spot) — the classic QLoRA card; runs 7B easily and a 13B model with sensible configuration; expect a meaningful step up in price, but it is the tier most serious builders target.
- Multi-GPU or larger cards (advanced) — needed once you want to train substantially bigger models; this is where a dual-card build earns its keep. Our step-by-step dual-GPU training rig build covers exactly how to assemble one.
For organisations that want a turnkey machine rather than a parts list, our overview of the AI training workstation for business ties the hardware decisions to real workloads.
Power, heat, and the NEPA factor
A point too often ignored: training runs take hours, sometimes a full working day. That is hours of your GPU pulling sustained power with no tolerance for interruption. A single NEPA outage at hour six wipes out the run and forces you to start again. An adequately sized UPS is not a luxury here — it is the difference between finishing a training job and repeating it. Budget for a UPS that can carry the rig long enough to either ride out a brief cut or shut down cleanly, and make sure your cooling can cope with hours of full load in a warm room.
Frequently Asked Questions
Is QLoRA worse than full fine-tuning? For some demanding tasks, full fine-tuning still has a small edge in raw capability. But for the vast majority of practical adaptation — teaching a model your tone, your domain, your formats — LoRA and QLoRA are excellent, and the VRAM savings are so transformative that the trade-off is almost always worth it.
Can I really fine-tune a 7B model on one consumer GPU? Yes. That is precisely what QLoRA was designed to enable. A single 24 GB card handles a 7B model comfortably, and even a 16 GB card can manage it with careful settings. This was simply impossible with full fine-tuning on the same hardware.
Do I need to keep the base model and the adapter together? The adapter is trained to sit on top of a specific base model, so you keep both. The upside is that the base model stays unchanged — you can store many small adapters, each for a different task, and load whichever one you need on top of the same base.
The One Thing to Remember
The reason LoRA and QLoRA are revolutionary is not that they make models smaller — it is that they remove the gradient and optimiser-state memory that made full fine-tuning so VRAM-hungry, and QLoRA then quantises the frozen base weights on top. Freeze most of the model, train a tiny adapter, and a job that once needed a server now fits on a single desktop card.
Ready to build a machine that can actually train models? Use our configurator to spec a workstation around the VRAM tier you need, or contact our team and we will help you match the hardware to your exact workload and budget.