AIINFRA 201: Model Adaptation — Fine-Tuning & Quantization — Syllabus

Course description

This course teaches learners how to adapt open-weight large language models to specialized tasks through supervised fine-tuning, parameter-efficient methods (LoRA, QLoRA), and preference tuning (DPO), all runnable on a single consumer or cloud GPU. Students then compress their adapted models through quantization (GGUF, AWQ, GPTQ, FP8) to prepare them for efficient, low-cost serving. The course emphasizes hands-on decision-making — when to fine-tune versus prompt or retrieve — and culminates in a capstone that packages a fine-tuned, quantized model ready for deployment.

Contact hours

54 contact hours (3-unit equivalent), delivered over a 16-week term, approximately 3.4 hours/week. This is a non-credit course on the CDCP Certificate of Completion pathway.

Prerequisite

AIINFRA 200 (Production Inference Serving; assumes Python and container fundamentals from earlier courses).

Course Learning Outcomes (CLOs)

CLO1: Evaluate a task and choose correctly among prompt engineering, retrieval-augmented generation, and fine-tuning — justifying the choice by cost, data availability, and hardware constraints.
CLO2: Build, clean, and format instruction/chat datasets and run supervised and parameter-efficient fine-tuning (LoRA, QLoRA, PEFT) on a single GPU.
CLO3: Apply preference tuning with DPO and track experiments while avoiding overfitting and catastrophic forgetting.
CLO4: Quantize models into GGUF, AWQ, GPTQ, and FP8 formats and analyze the precision, quality, and memory tradeoffs.
CLO5: Assess licensing and ethical constraints on model weights and data, and package an adapted, quantized model for serving.

Weekly schedule

Week	Topic
01	Adapt, Prompt, or Retrieve: The Model Customization Decision
02	Open Models, Licenses, and the Fine-Tuning Toolchain Setup
03	Dataset Design and Data Preparation for Fine-Tuning
04	Instruction and Chat Formatting, Tokenization, and Templates
05	Supervised Fine-Tuning Fundamentals with Hugging Face TRL
06	Parameter-Efficient Fine-Tuning: LoRA and PEFT
07	QLoRA and Training on Consumer GPUs with Unsloth
08	Axolotl, Config-Driven Training, and Experiment Tracking (Midterm)
09	Preference Tuning: DPO in Practice (with a Look at GRPO)
10	Evaluating Fine-Tuned Models and Detecting Overfitting
11	Catastrophic Forgetting, Adapter Merging, and Cloud Training
12	Quantization Foundations: Precision, GGUF, and llama.cpp
13	Advanced Quantization: AWQ, GPTQ, FP8, and KV-Cache Quantization
14	Knowledge Distillation and Quality-Size Tradeoff Analysis
15	Serving Adapted and Quantized Models — Handoff to AIINFRA 200
16	Capstone Project & Course Review (Capstone)

Grading

Component	Weight
Labs	40%
Discussions	10%
Weekly Quizzes	15%
Midterm	15%
Final Capstone	20%

Credit/No-Credit grading. 70% overall is required to pass.

Course policies

Academic integrity: Students must submit their own work. Collaboration on concepts is encouraged, but submitted code, datasets, and written analysis must reflect the student's own understanding. Fabricated results or misrepresented model outputs are treated as academic dishonesty under LACCD student conduct policy. Late work: Assignments submitted after the deadline may incur a grade reduction unless an extension is arranged in advance with the instructor. Extenuating circumstances are handled on a case-by-case basis — contact the instructor as soon as possible. Responsible use of AI: This course centers on building and adapting AI models, so use of AI tools for coding assistance, debugging, and drafting is permitted and encouraged where it supports learning. Students must be able to explain and defend any submitted work, including AI-assisted portions, and must not submit AI output as-is without understanding or verifying it. Accessibility: This course complies with ADA and Section 508 accessibility standards. Students needing accommodations should contact the campus Disabled Students Programs and Services (DSPS) office and notify the instructor early in the term so materials and labs can be adapted appropriately.

Tools & materials

All tools and materials are free or available on a free tier.

Fine-tuning libraries: Hugging Face Transformers, PEFT, TRL
Accelerated training: Unsloth, Axolotl, bitsandbytes
Quantization: llama.cpp/GGUF, AutoAWQ, GPTQModel, llm-compressor (FP8)
Experiment tracking: Weights & Biases, TensorBoard, Hugging Face Hub
Compute platforms: Google Colab, Kaggle Notebooks, RunPod, Lambda Cloud
Datasets & models: Hugging Face Datasets, open weights (Llama/Qwen/Mistral/Gemma)
Evaluation: lm-evaluation-harness, LLM-as-judge, held-out test sets
Serving handoff: Ollama, vLLM, LM Studio, text-generation-inference

Week 01 · Adapt, Prompt, or Retrieve: The Model Customization Decision

Notion ID: 392c08fd-0278-8119-94db-db3da8f865fe Notion URL: https://app.notion.com/p/392c08fd0278811994dbdb3da8f865fe