What is LoRA (Low-Rank Adaptation)

 

LoRA is a parameter-efficient fine-tuning technique used to adapt large language models (LLMs) like LLaMA, GPT, etc., to new tasks without retraining the entire model.

Instead of updating all the billions of parameters, LoRA:

• Freezes the original model weights (keeps them unchanged)

• Inserts small trainable low-rank matrices into certain layers (usually attention layers)

• Only trains these small matrices, which are much smaller than the full model

---

⚙️ How LoRA Works (Simplified)

Imagine an LLM has a large weight matrix W (like 4096×4096).

Normally, fine-tuning means updating all entries in W → which is huge.

With LoRA:

1. Keep W frozen.

2. Add two small matrices:

   • A (size 4096×r)

   • B (size r×4096) — where r is small (like 8 or 16)

3. Train only A and B.

4. At inference time, the effective weight becomes:

   W' = W + A × B
   

This drastically reduces the number of trainable parameters.

---

📊 Why LoRA is Useful

Aspect	Full Fine-Tune	LoRA Fine-Tune
Parameters updated	All (billions)	Few million (<<1%)
GPU memory need	Very high	Very low
Training speed	Slow	Fast
Sharing	Must share full model	Just share small LoRA weights

This makes LoRA ideal when:

• You want to customize a big model on a small dataset

• You have limited GPU resources

• You want to train multiple variants of the same base model

---

📦 Common Uses

• Domain-specific tuning (medical, legal, finance text)

• Instruction tuning or chat-like behavior

• Personalizing models for specific companies or users

• Combining with PEFT (Parameter-Efficient Fine-Tuning) frameworks like:

  • 🤗 Hugging Face PEFT

  • 🤖 bitsandbytes

  • 🦙 LLaMA + LoRA (common combo)

---

📝 Summary

LoRA = a lightweight way to fine-tune large models by training only tiny "adapter" layers (low-rank matrices) while keeping original weights frozen.
It dramatically reduces cost, time, and storage needs for customizing LLMs.

ValueError: Some modules are dispatched on the CPU or the disk. Make sure you have enough GPU RAM to fit the quantized model. If you want to dispatch the model on the CPU or the disk while keeping these modules in 32-bit, you need to set llm_int8_enable_fp32_cpu_offload=True and pass a custom device_map to from_pretrained. Check https://huggingface.co/docs/transformers/main/en/main_classes/quantization#offload-between-cpu-and-gpu for more details.

 

hat error usually pops up when you try to load a quantized model (like 4-bit QLoRA) with Hugging Face’s transformers and bitsandbytes, but your GPU doesn’t have enough VRAM to hold all the model layers. By default, Hugging Face tries to put everything on the GPU, and when it can’t, it errors out.

---

🔹 Why the Error Happens

• The model is too large for your GPU (e.g., a 13B model on 8GB VRAM).

• Some layers got “dispatched” (offloaded) onto CPU or disk.

• You didn’t explicitly allow CPU offloading, so transformers raised a ValueError.

---

🔹 Solutions

✅ Option 1: Enable CPU Offload (Recommended if low VRAM)

You can allow the model to split between GPU + CPU:

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

model_id = "mistralai/Mistral-7B-Instruct-v0.2"

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype="bfloat16"
)

tokenizer = AutoTokenizer.from_pretrained(model_id)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    quantization_config=bnb_config,
    device_map="auto",  # automatically spread across GPU + CPU
    llm_int8_enable_fp32_cpu_offload=True  # ✅ allows CPU fallback in FP32
)

---

✅ Option 2: Force Specific Device Map

If you want more control, specify which parts go where:

device_map = {
    "transformer.wte": "cuda:0",   # embeddings on GPU
    "transformer.h": "cpu",        # some layers on CPU
    "lm_head": "cuda:0"
}

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    quantization_config=bnb_config,
    device_map=device_map,
    llm_int8_enable_fp32_cpu_offload=True
)

---

✅ Option 3: Use a Smaller Model

If you only have ~8–12 GB VRAM, try:

• Mistral-7B-Instruct-v0.1 instead of v0.2

• Or even smaller models (LLaMA-2-7B, Gemma-2B, etc.).

---

✅ Option 4: Use accelerate for Better Device Placement

pip install accelerate

Then run:

from accelerate import init_empty_weights, load_checkpoint_and_dispatch

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    quantization_config=bnb_config,
    device_map="auto"
)

This lets accelerate decide where to put layers across GPU/CPU/Disk.

---

🔹 Key Takeaway

• If GPU VRAM < model size → must offload to CPU/disk.

• Add llm_int8_enable_fp32_cpu_offload=True and/or device_map="auto".

• Or use a smaller model to fit fully in GPU.

LoRA vs QLoRA

 

🔹 1. LoRA (Low-Rank Adaptation)

LoRA is a parameter-efficient fine-tuning method for large language models (LLMs).

🧠 The Core Idea

• Instead of updating all parameters of a huge LLM (billions of weights), LoRA inserts small trainable matrices (low-rank adapters) into certain layers (usually attention and/or feed-forward layers).

• During fine-tuning:

  • Base model weights stay frozen (unchanged).

  • Only the small adapter weights are trained.

This massively reduces:

• Memory usage 💾

• Compute cost ⚡

• Training time ⏱️

---

🔹 LoRA Example

If a weight matrix is W (say 4096 × 4096), instead of fine-tuning all ~16M parameters, LoRA trains two small matrices:

• A (4096 × r) and B (r × 4096), where r is the rank (say 8 or 16).

• The effective update is:

  W' = W + A × B
  

So you only train a few thousand parameters instead of millions.

---

🔹 2. QLoRA (Quantized LoRA)

QLoRA takes LoRA one step further by adding quantization.

🧠 The Core Idea

• Quantization = Compress model weights into fewer bits (e.g., 16-bit → 4-bit).
  This saves GPU memory and makes training possible on smaller hardware.

• QLoRA fine-tunes the quantized model with LoRA adapters on top.

So:

1. Base model → 4-bit quantized (efficient storage + inference).

2. Train only LoRA adapters (small rank matrices).

3. Combine for final fine-tuned model.

---

🔹 Why QLoRA is Powerful

• You can fine-tune 13B+ parameter models on a single consumer GPU (24GB VRAM).

• Example: Guanaco, Alpaca, Vicuna fine-tunes often use QLoRA.

• Enables democratization → people without supercomputers can fine-tune LLMs.

---

🔹 LoRA vs QLoRA (Quick Comparison)

Feature	LoRA	QLoRA
Base Model	Full precision (16-bit/32-bit)	Quantized (4-bit/8-bit)
Memory Usage	Medium (needs decent GPU)	Very low (fits big models on consumer GPUs)
Training	Adapter training only	Adapter training only (on quantized model)
Speed	Fast	Even faster (smaller memory)
Trade-off	Slightly more accurate	Small accuracy drop due to quantization

---

🔹 Visual Analogy

• LoRA = Adding small “adjustment knobs” to a giant machine, instead of rebuilding the whole machine.

• QLoRA = Compressing the giant machine first, then adding the small adjustment knobs.

---

✅ In practice:

• Use LoRA if you have strong GPU resources.

• Use QLoRA if you want to fine-tune big models (7B–65B) on consumer GPUs (like RTX 3090, 4090, A100 40GB).