Fine-Tuning and Analysis of 1-Bit LLMs with QLoRa Adapters and other quantizations

This project explores and compares quantization techniques for efficiently fine-tuning large language models, with a focus on QLoRA and BitNet. The base model is GPT-2, and scalability is tested on a LLaMA-7B model. The goal is to reduce memory usage and latency without significant performance loss.

---

🔍 Project Objectives

• Evaluate various fine-tuning strategies:
  • Full-precision baseline
  • 8-bit and 4-bit quantization
  • QLoRA (4-bit LoRA adaptation)
  • BitNet (1.58-bit ternary compression)
  • Combined QLoRA + BitNet
• Measure:
  • Perplexity
  • Training time
  • Inference latency
  • Peak GPU memory usage
  • Trainable vs. total parameters
• Scale the best approach to a 7B-parameter model
• Deploy and benchmark on Google Cloud (T4 GPU)

---

📦 Dataset

• CodeSearchNet (Python subset)
  Source: Papers With Code
  Contains millions of function-documentation pairs for supervised code modeling.

---

🧪 Techniques Used

Method	Quantization	Trainable Params	Notes
Baseline Fine-Tune	None (FP32)	All	High accuracy, slow training
8-bit / 4-bit	8 / 4-bit	All	Reduced memory footprint
QLoRA	4-bit (NF4)	LoRA adapters	Efficient fine-tuning
BitNet	1.58-bit	Ternary linear	Custom CUDA; inference only
QLoRA + BitNet	1.58-bit	LoRA only	Best for deployment

---

📈 Metrics Collected

• Perplexity
• Training Time
• GPU Peak Memory Usage
• Inference Latency
• Trainable Parameters
• Total Parameters

---

🛠️ Setup Instructions

# Create environment
python -m  venv .venv
source .venv/bin/activate

# Install requirements
pip install -r requirements.txt