Learn how to fine-tune a large language model (LLM) using QLoRA (Quantized Low-rank Adaptation) on a single GPU in this comprehensive 37-minute video tutorial. Explore the four key ingredients of QLoRA: 4-bit NormalFloat, Double Quantization, Paged Optimizer, and LoRA. Follow along with example Python code to train a custom YouTube comment responder using Mistral-7b-Instruct. Gain insights into quantization techniques, computational efficiency, and practical implementation. Access additional resources including a series playlist, related videos, blog post, Colab notebook, GitHub repository, and Hugging Face model and dataset links for further learning and experimentation.
Overview
Syllabus
Intro -
Fine-tuning recap -
LLMs are computationally expensive -
What is Quantization? -
4 Ingredients of QLoRA -
Ingredient 1: 4-bit NormalFloat -
Ingredient 2: Double Quantization -
Ingredient 3: Paged Optimizer -
Ingredient 4: LoRA -
Bringing it all together -
Example code: Fine-tuning Mistral-7b-Instruct for YT Comments -
What's Next? -
Taught by
Shaw Talebi
Related Courses
-
Fine-Tuning LLM Models - Generative AI Course
-
Improving Accuracy of LLM Applications
-
Full Fine-Tuning vs LoRA and QLoRA - Comparison and Best Practices
-
QLoRA: Efficient Training of Large Language Models Using Quantization and Low-Rank Adaptation
-
Understanding 4-bit Quantization and QLoRA - Memory Efficient Fine-tuning of LLMs
-
Local LLM Fine-tuning on Mac (M1 16GB) Using QLoRA and MLX
