Course Staff

Dr. Murari Mandal
Instructor
Umakanta Mahrana
Instructor
Debdeep Sanyal
Instructor
Aakash Sen Sharma
Instructor
Arka Mukherjee
Instructor

Content

What is this course about?

Language models serve as the cornerstone of modern natural language processing (NLP) applications and open up a new paradigm of having a single general purpose system address a range of downstream tasks. As the field of artificial intelligence (AI), machine learning (ML), and NLP continues to grow, possessing a deep understanding of language models and generative ai in general becomes essential for scientists and engineers alike. This course is designed to provide students with a comprehensive understanding of language models by walking them through the entire process of how a language model is developed and optimized. Unlike most of the courses, this course will include a lot of coding the concepts along you learn along the way, which includes developing an entire language model and from scratch and optimizing it in various aspects.

Prerequisites

Note that this is not a graded class. The purpose of this class is to introduce the mechanistics and system level optimizations of large generative models to students which are not available as distinct courses on the internet.


Coursework

Assignments

Coding assignments, related to each topic, will be provided when required and discussed in class. While ungraded, these assignments are crucial for solidifying your understanding. For any assistance, students are welcome to contact any instructor. All of the codes discussed in the classes will be available on github. Here's a list of questions you can try out before joining the class. Feel free to refer to ChatGPT or any AI to help solve and understand the questions.

Topics to be covered

Module 1: Introduction and Foundations
  • Course Introduction
  • A general history of the progression of LLMs
  • Natural Language Processing (NLP)
    • Statistical Language Modeling
    • Advanced Smoothing and Evaluation Methods
    • Word Representation: Word2vec
Module 2: Neural Language Modeling
  • Neural Networks (NN)
  • Convolutional Neural Networks (CNN)
  • Recurrent Neural Networks (RNN)
  • Long Short-Term Memory Networks (LSTM)
  • Sequence-to-Sequence Models
  • Backpropagation
Module 3: Training Neural Network Models
  • Attention Mechanisms
  • Optimizers
  • Residual Connections
  • Batch Normalization / Layer Normalization
Module 4: Transformers
  • Introduction to Transformers
  • "Attention is All You Need" Paper
  • Tokenizers (from scratch and using HuggingFace libraries)
Module 5: LLM Training
  • Pre Training
    • Data Preparation
    • Scaling Laws
  • Fine-tuning (TULU)
  • In-Context Learning
  • Preference Optimization
    • RLHF (Reinforcement Learning from Human Feedback)
    • DPO (Direct Preference Optimization)
    • PPO (Proximal Policy Optimization)
    • GRPO (Group Relative Policy Optimization)
  • Distributed Training
    • Fully Sharded Data Parallel
    • DeepSpeed
    • Accelerate
Module 6: LLM Optimization
  • Inference Optimization
    • Batching
    • Quantization
      • FP32, FP16, int8, NF4
      • GGUF
      • llama.cpp
    • KV Caching
    • torch.compile()
    • FlashAttention
  • Attention Optimization (L Serve, Top-k, Top-Theta)
  • Parameter-Efficient Fine-Tuning (LoRA, QLoRA, etc.)
  • Distillation
Module 7: Reasoning with LLMs
  • Search Techniques
  • Test-Time Scaling
  • Reinforcement Learning (RL) Overview
  • Chain-of-Thought (CoT) Reasoning
  • Related Research Papers
Module 8: Advanced LLM Architectures
  • Multimodal LLMs
  • Mixture of Experts (MoE)
  • Mamba
  • Diffusion LLMs
  • Large Concept Models (LCM)
Module 9: Retrieval-Augmented Generation (RAG)
  • Indexing
  • Retrieval
  • Generation
  • Query Translation
    • Multi Query
    • RAG Fusion
    • Decomposition
    • HyDE
  • Routing
  • Query Construction
  • Indexing
  • Building a RAG application
Module 10: Diffusion Models
  • Autoencoders
  • Variational Autoencoders (VAEs)
  • UNets
  • Diffusion Transformers
  • Diffusion Process via Markov Chains
  • Timestep Schedulers (cosine, linear, beta, etc.)
  • Classifier-Free Guidance
  • Vision Language Models
  • Cross attention based guidance

Announcements 📢

The second lecture is out! Slides for the second class will be made available soon.

Schedule

# Date Description Course Materials Exercises Video Link
1 19th March 2025 Introduction to Language Models: Scope and Overview Introduction to Language Models - Scope and Overview Assignment 1 LLM | Introduction to Language Models: Scope and Overview | Lecture 01
2 23rd March 2025 Introduction Neural Networks - Part 1 TBD TBD LLM | Introduction to Neural Networks: Part 1 | Lecture 02
3 30th March 2025 Advanced Language Modeling with Recurrent Nets Advanced Language Modeling with Recurrent Networks Resources on RNNs and LSTMs NA