Orientation

Large Language Models : A Hands-on Approach

CCE, IISc

Course Overview

Duration: ~12 Weeks
Format: 2-4 sessions per week (Tue-Fri, 7:00–8:30 PM IST)
Style: Lectures, demos and labs
Assessment: Quizzes, assignments, and final project/presentation
Credits: 3 + 1 (Credits added in ABC)

Why This Course?

LLMs have changed how we build intelligent applications.

This course aims to bridge the gap between:

Using LLMs via APIs -> Engineering LLM systems

Learning Outcomes

Understand LLM architectures
Understanfd GPU programming and optimization techniques
Optimize inference for cost and latency
Fine-tune models for specific domains and tasks
Build robust LLM-powered applications (RAG, agents, tool use)
Evaluate, debug, and improve model performance
Leverage open-source ecosystems for LLM engineering

What This Course Will NOT Cover

Mathematical foundations
Pretraining LLMs from scratch
State of the art in research
Data engineering pipelines
Ethics and societal impacts

Prerequisites

Programming: Python proficiency
ML Basics: Understanding of neural networks and deep learning
Helpful: Familiarity with PyTorch, basic NLP concepts

Tools and Resources

Software Stack

Python
PyTorch
HuggingFace Ecosystem(models, libraries and datasets)

Compute Stack

Option	Cost	Notes
Google Colab	Free	Limited GPU, good for experiments
GCP Free Credits	Free	$300 for new accounts
Colab Pro+	Paid	More GPU time, better GPUs
RunPod	Paid	Flexible, longer runs
Lightning AI	Paid	Easy setup, good UX

Free options suffice for most of our course needs

Evaluation Criteria

Component	Weight	Notes
Quizzes	20%	5 short in-class quizzes
Assignments	40%	2 assignments (May and June)
Final Project/Presentation	40%	June 3rd week onwards

–

ABC

Academic Bank of Credits
Credits will be added to your ABC account after successful course completion
More information will be shared during the course

Final Project Option

Individual or group project (max 2).

Build an end-to-end LLM application:

Model fine-tuning
Inference optimization
LLM Application (RAG, Agents)
Open source contribution
Other ideas welcome too

Deliverables: Documentation + Demo / Presentation

Presentation Option

Topic: Novel concept not covered in class
Format: Research paper or case study
Deliverables: Slides or summary document
Evaluation: Depth of understanding, presentation quality
Duration: 10 minutes, 8-10 slides

Course Roadmap (1/2)

Module	Topic
1	LLM Foundations - Transformers, GPT-2, Modern Architectures, MoE, OSS Models
2	GPUs - Architecture and Programming, Multi-GPU Parallelism,
3	Inference - Sampling, KV Caching, Quantization, Speculative Decoding, Model Serving
4	RAG & Agents - RAG Fundamentals, ReAct, Tools, Protocols, Agents, Finetuning
5	Evaluation - Moderl Evaluations, Benchmarks, LLM-as-a-Judge

Course Roadmap (2/2)

Module	Topic
6	Fine-Tuning - SFT, PEFT (LoRA, QLoRA etc),RLHF (DPO, GRPO etc), Distillation
7	Reasoning - Chain-of-Thought (CoT), Test Time Scaling, Finetuning
8	Multimodal - Multimodal Architectures, Finetuning
9	Edge Deployment - Edge architectures, Optimization and Deployment

Communication Channels

In-class: Ask questions during sessions
Email: For updates and announcements
Forum: Discord discussions, doubts, and sharing resources (Link to be shared)

Making the Best of This Course

Participate actively: Ask questions in class, engage in discussions
Practice hands-on: Run the code, experiment with parameters
Use AI tools: Leverage AI assistants for coding and learning
Stay connected: Use forum for async discussions