Fine‑Tuning Transformers vs LoRA vs QLoRA 2024 – What You Need to Know
Hey folks, Nick Creighton here. If you’ve been listening to the latest Build Log episode you know I’m all about shipping code that actually moves the needle. In this post I’m taking the audio‑first conversation we just had and turning it into a step‑by‑step guide you can read, bookmark, and act on.
Why This Comparison Matters in 2024
Just a year ago the default way to adapt a large language model (LLM) was to re‑train every single parameter. That meant pulling a ggml or torch checkpoint, slamming a GPU farm together, and waiting hours (or days) for a new .bin file to appear. The result? A massive artifact that cost you in storage, latency, and maintenance.
Fast‑forward to today: LoRA and its cousin QLoRA have become the de‑facto tools for most production teams. They let you add a tiny adapter to a frozen model, keep the base weights untouched, and ship a few megabytes of delta‑weights instead of gigabytes of new model. The upside is immediate – lower cost, faster iteration, and a simpler deployment pipeline.
Below you’ll find the practical, no‑fluff details you need to decide which approach fits your project, how to set it up, and which pitfalls to avoid.
1. Traditional Full‑Model Fine‑Tuning – The “Old Guard”
When I first started experimenting with LLMs, the workflow looked like this:
- Pick a base model (e.g., Llama‑2‑7B).
- Load the checkpoint into PyTorch.
- Create a Trainer (or accelerate) with gradient_accumulation_steps to fit the model in memory.
- Run trainer.train() for N epochs on your domain data.
- Save the resulting checkpoint – often a 10‑15 GB file.
That’s all well and good if you have a dedicated server farm and a budget that looks like a startup’s seed round. But for most indie developers, SaaS founders, or data‑science hobbyists, the cost‑to‑value ratio is terrible.
When It Still Makes Sense
- You need to modify every layer – e.g., adding new token embeddings, changing the tokenizer, or altering the model’s architecture.
- Your downstream task is extremely sensitive to subtle weight changes (e.g., medical diagnostics where you need the absolute best performance).
- You have a budget for GPU hours (think 2–4 A100s for a week).
2. LoRA – Low‑Rank Adaptation, The Tiny Cheat Sheet
LoRA is essentially a matrix factorisation trick. Instead of updating the full weight matrix W ∈ ℝ^{d×k}, you freeze W and learn two much smaller matrices A ∈ ℝ^{d×r} and B ∈ ℝ^{r×k} such that ΔW = A·B. The rank r is typically 4–64 – orders of magnitude smaller than d or k.
Why It Works
- Parameter Efficiency: You only store A and B. For a 7‑B model, a LoRA adapter can be as small as 5–30 MB.
- Speed: The forward pass adds a cheap low‑rank matrix multiplication, barely affecting latency.
- Reversibility: Because the base model stays frozen, you can swap adapters on the fly (multi‑tasking becomes trivial).
Getting Started with LoRA
- Install the right libraries – I use peft (🤗 PEFT) and accelerate. pip install peft accelerate transformers
- Pick a base model (e.g., meta-llama/Llama-2-7b-chat-hf) and load it with torch_dtype=torch.float16 to keep VRAM low.
- Configure the LoRA adapter. from peft import LoraConfig, get_peft_model
lora_cfg = LoraConfig(
r=32, # rank
lora_alpha=64,
target_modules=["q_proj", "v_proj"], # typical for Llama
lora_dropout=0.05,
bias="none",
)
model = get_peft_model(base_model, lora_cfg)
- Prepare your dataset – a simple .jsonl with {"prompt": "...", "completion": "..."} works fine.
- Train using Trainer or accelerate launch. For a 1 GB dataset, 3‑4 epochs on a single RTX 4090 take ~15 minutes.
- Save only the adapter. model.save_pretrained("my_lora_adapter")
Actionable Tip #1 – Keep the Rank Low, Then Scale
If you’re not sure about the optimal rank, start with r=8. Evaluate on a validation set. If performance plateaus, bump to 16 or 32. The memory overhead grows linearly with r, so you’ll instantly see the trade‑off.
3. QLoRA – Quantized LoRA, The Sweet Spot for 2024
QLoRA builds on LoRA by quantising the base model to 4‑bits (or 8‑bits) using bitsandbytes while still training the adapter in full‑precision. The result: you can fine‑tune a 70‑B model on a single 24 GB GPU!
Key Benefits
- GPU Efficiency: 4‑bit quantisation reduces VRAM by ~80 %.
- Training Speed: Less data movement = higher throughput.
Quality Retention: Empirical studies (including my own benchmarks) show Setting Up QLoRA
Install bitsandbytes (CUDA‑compatible version).
pip install bitsandbytesLoad the model with load_in_4bit=True and set bnb_4bit_compute_dtype=torch.float16.
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained(
"meta-llama/Llama-2-70b-chat-hf",
device_map="auto",
load_in_4bit=True,
bnb_4bit_compute_dtype=torch.float16,
bnb_4bit_quant_type="nf4"
)
- Apply the same LoraConfig as before – QLoRA works with any PEFT adapter.
- Train exactly as you would with LoRA. The only difference is the underlying model is quantised.
- When you export, you can either keep the 4‑bit checkpoint (ideal for inference) or de‑quantise for a higher‑precision fallback.
Actionable Tip #2 – Use nf4 Quantisation for Better Stability
Bitsandbytes offers two 4‑bit schemes: fp4 and nf4. nf4 (normalised float‑4) tends to preserve the distribution of weights better, which translates to less “catastrophic forgetting” during LoRA training. If you hit sudden spikes in loss, switch to nf4.
4. Decision Matrix – Which Tool for Which Job?
Criterion
Full‑Model FT
LoRA
QLoRA
GPU Budget
Multiple A100‑40G or V100‑32G
Single RTX 4090 / A6000
Single 24 GB GPU (RTX 4090, A6000)
Model Size
Up to ~13 B comfortably
Any size (adapter tiny)
Up to 70 B (quantised)
Deployment Complexity
High – new artifact, versioning
Low – swap adapters
Low – same as LoRA, but smaller runtime
Performance Gap vs Full‑FT
0 % (baseline)
~2–5 % on average
~1–3 % on average
Use‑Case Fit
Token‑embedding changes, architecture tweaks
Domain‑specific chat, classification, summarisation
Large‑scale embeddings, retrieval‑augmented generation, heavy traffic services
- End‑to‑End Workflow You Can Copy‑Paste
Below is a minimal script that works for both LoRA and QLoRA. Swap the load_in_4bit flag to toggle.
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, Trainer, TrainingArguments
from peft import LoraConfig, get_peft_model
from datasets import load_dataset
-------------------------------------------------
1️⃣ Load tokenizer & model
-------------------------------------------------
model_name = "meta-llama/Llama-2-7b-chat-hf"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
model_name,
device_map="auto",
torch_dtype=torch.float16,
# Uncomment the next three lines for QLoRA
# load_in_4bit=True,
# bnb_4bit_compute_dtype=torch.float16,
# bnb_4bit_quant_type="nf4",
)
-------------------------------------------------
2️⃣ Attach LoRA adapter
-------------------------------------------------
lora_cfg = LoraConfig(
r=32,
lora_alpha=64,
target_modules=["q_proj", "v_proj"],
lora_dropout=0.05,
bias="none",
)
model = get_peft_model(model, lora_cfg)
-------------------------------------------------
3️⃣ Prepare data (simple jsonl)
-------------------------------------------------
data = load_dataset("json", data_files={"train": "train.jsonl", "valid": "valid.jsonl"})
def tokenize_fn(example):
tokens = tokenizer(example["prompt"], truncation=True, max_length=512)
tokens["labels"] = tokenizer(example["completion"], truncation=True, max_length=512)["input_ids"]
return tokens
tokenized = data.map(tokenize_fn, batched=True, remove_columns=["prompt", "completion"])
-------------------------------------------------
4️⃣ Training arguments
-------------------------------------------------
training_args = TrainingArguments(
output_dir="outputs",
per_device_train_batch_size=4,
gradient_accumulation_steps=4,
num_train_epochs=3,
learning_rate=2e-4,
fp16=True,
logging_steps=20,
save_steps=200,
evaluation_strategy="steps",
eval_steps=100,
load_best_model_at_end=True,
)
-------------------------------------------------
5️⃣ Trainer & launch
-------------------------------------------------
trainer = Trainer(
model=model,
args=training_args,
train_dataset=tokenized["train"],
eval_dataset=tokenized["valid"],
)
trainer.train()
model.save_pretrained("my_adapter")
tokenizer.save_pretrained("my_adapter")
print("✅ Training complete – adapter saved!")
Actionable Tip #3 – Use Gradient Accumulation to Fit Bigger Batches
Even on a 24 GB card you can simulate a batch size of 32–64 by setting per_device_train_batch_size=4 and gradient_accumulation_steps=8. Larger effective batches improve stability, especially with low‑rank adapters.
6. Cost & Speed Benchmarks (My Real‑World Numbers)
- Full‑Model FT (7 B) – 1 GPU × A100‑40G, 2 hours, $6.80 (Azure pay‑as‑you‑go).
- LoRA (7 B) – 1 GPU × RTX 4090, 15 minutes, $0.12.
- QLoRA (70 B) – 1 GPU × RTX 4090, 45 minutes, $0.35.
Those numbers assume standard on‑demand pricing and a modest 10 GB dataset. The takeaway: you can get production‑grade results for pennies.
7. Common Pitfalls & How to Dodge Them
- Adapter Size Too Large – If your r is > 128 you’ll lose the memory advantage. Keep an eye on torch.cuda.memory_allocated().
- Forgetting to Set torch_dtype – Mixing float32 and float16 leads to “CUDA out of memory” errors. Explicitly set torch_dtype=torch.float16 when loading the base.
- Validation Drift – Because the base model is frozen, any over‑fitting shows up quickly in the validation loss. Use early stopping (early_stopping_patience=2).
- Quantisation Instability – With QLoRA, occasionally the loss spikes after a few hundred steps. Reduce the learning rate to 1e-4 and add lora_dropout=0.1.
- Adapter Compatibility – When swapping adapters between models, ensure the target modules list matches the architecture (e.g., “q_proj” exists in both Llama‑2 and Mistral).
8. Real‑World Use Cases I’m Running Right Now
To prove the point, here are three production pipelines I’ve deployed on thirteen active websites:
- Customer‑Support Chatbot – LoRA on Llama‑2‑7B, r=16. Handles 5k QPS with ~40 ms latency on a single RTX 4090. Legal‑Document Summariser – Q
Adapted from an episode of Signal Notes. Listen on your favorite podcast app.