LLMs - PyTorch - HuggingFace - LangChain

What you are going to build in the projects (May be subject to changes)

• Project 1: Implementing from scratch The sparse attention mechanisms, SliGLU, RMSNorm, MoE, and Rope embedding in PyTorch
• Project 2: Fine-tuning an LLM with PPO vs DPO vs ORPO using the PEFT package.
• Project 3: Train an LLM in a distributed manner with the Accelerate package in AWS SageMaker with the Zero Redundancy Optimizer Strategy.
• Project 4: Fine-tuning a model with QLoRA to increase the context size.
• Project 5: Deploying a scalable LLM application API with streaming, KV-caching, Continuous batching, and text generation layer capabilities. 
• Project 6: Deploying an RAG application using LangChain, FastAPI, and LangServe.
  

6 Weeks of Intense Learning!

The Transformer Architecture (1 week)

The Transformer is the fundamental Neural Network architecture that enabled the evolution of Large Language Models as we know them now.

• The Self-Attention Mechanism
• The Multihead attention
• The encoder-decoder architecture
• The position embedding
• The layer-normalization
• The position-wise feed-forward network
• The cross-attention layer
• The language modeling head
  

Training LLMs to Follow Instruction (1 week)

GhatGPT, Claude, or Gemini are LLMs trained to follow human instructions. We are going to learn how those are trained from scratch:

• The Causal Language Modeling Pretraining Step
• The Supervised Learning Fine-Tuning Step
• The Reinforcement Learning Fine-Tuning Step
• Implementing those Steps with HuggingFace
  

How to Scale Model Training (1 week)

More than ever, we need efficient hardware to accelerate the training process. We are going to explore the strategy of distributing training computations across multiple GPUs for different parallelism strategies:

• CPU vs GPU vs TPU
• The GPU Architecture
• Distributed Training
• Data Parallelism
• Model Parallelism
• Zero Redundancy Optimizer Strategy
  

How to Fine-Tune LLMs (1 week)

Fine-tuning a model means we continue the training on a specialized dataset for a specialized learning task. We are going to look at the different strategies to fine-tune LLMs:

• The different fine-tuning learning tasks
• Catastrophic forgetting
• LoRA Adapters
• QLoRA

How to Deploy LLMs (1 week)

The most important part of a machine learning model development is the deployment! A model that is not in production is a model that is costing money instead of generating money for the company. We are going to explore the different strategies to deploy LLMs:

• The Deployment Strategies
• Multi-LoRA
• The Text Generation Layer
• Streaming Applications
• Continuous Batching
• KV-Caching
• The Paged-Attention and vLLM

Building the Application Layer (1 week)

A deployed LLM on its own is not really useful. We are going to look at how we can build an agentic application on top of the model with LangChain:

• Implementing a Retriever Augmented Generation (RAG) pipeline with LangChain
• Optimizing the RAG pipeline
• Serving the pipeline with LangServe and FastAPI