NCP-GENL Generative AI & LLMs Professional Study Plan¶

8-Week Intensive Study Schedule¶

Phase 1: Foundation Building (Weeks 1-2)¶

Week 1: LLM Architecture and Transformer Fundamentals¶

Focus: Core architecture concepts and model families

Day 1-2: Transformer Deep Dive¶

• Review transformer architecture - attention, FFN, normalization
• Study self-attention mechanism and multi-head attention math
• Understand positional encoding - sinusoidal, learned, RoPE
• Read about pre-norm vs post-norm layer normalization
• Reference: NVIDIA NeMo GPT Training Guide

Day 3-4: Model Families and Tokenization¶

• Study GPT-style decoder-only architectures
• Learn LLaMA design choices - RoPE, SwiGLU, GQA
• Understand Mixture of Experts (MoE) architecture and routing
• Study tokenization - BPE, SentencePiece, vocabulary tradeoffs
• Lab: Explore different tokenizers using HuggingFace tokenizers library

Day 5-7: Scaling Laws and Model Selection¶

• Study Chinchilla scaling laws and compute-optimal training
• Learn about training FLOPS estimation
• Understand model size vs data size tradeoffs
• Review multi-modal architectures (vision-language models)
• Practice: Calculate memory requirements for different model sizes

Week 2: NeMo Framework and Training Basics¶

Focus: Setting up NeMo and understanding distributed training

Day 1-2: NeMo Framework Setup¶

• Install NeMo Framework using NGC containers
• Explore NeMo configuration system (Hydra/OmegaConf)
• Understand NeMo model checkpoints and formats
• Run a basic NeMo training example
• Reference: NeMo Framework Documentation

Day 3-4: Distributed Training Fundamentals¶

• Study data parallelism and gradient synchronization
• Learn tensor parallelism (Megatron-style column/row splits)
• Understand pipeline parallelism and micro-batching
• Study 3D parallelism combinations
• Reference: NeMo Parallelism Guide

Day 5-7: Training Best Practices¶

• Study mixed precision training - BF16, FP16, loss scaling
• Learn gradient accumulation and effective batch size
• Understand learning rate schedules - warmup, cosine decay
• Practice gradient clipping and training stability techniques
• Lab: Configure and run a multi-GPU training job with NeMo

Phase 2: Fine-Tuning and Inference (Weeks 3-5)¶

Week 3: Fine-Tuning Methods¶

Day 1-2: PEFT Methods¶

• Study LoRA - low-rank matrix decomposition in attention layers
• Learn QLoRA - quantized base model with LoRA adapters
• Understand P-tuning - soft prompt optimization
• Compare adapter methods and their memory profiles
• Reference: NeMo PEFT Guide

Day 3-4: SFT and RLHF¶

• Study Supervised Fine-Tuning (SFT) data formats and pipelines
• Learn reward model training from preference data
• Understand PPO for RLHF training
• Study Direct Preference Optimization (DPO) as alternative
• Lab: Fine-tune a model with LoRA using NeMo Framework

Day 5-7: Data Preparation¶

• Study NeMo Data Curator for data processing
• Learn data quality filtering and deduplication
• Understand instruction formatting and chat templates
• Practice data curation for fine-tuning workflows
• Practice: Prepare a custom dataset for fine-tuning

Week 4: Inference Optimization with TensorRT-LLM¶

Day 1-2: TensorRT-LLM Fundamentals¶

• Install and configure TensorRT-LLM
• Understand model compilation to TensorRT engines
• Learn the TensorRT-LLM Python API
• Study supported model architectures
• Reference: TensorRT-LLM Quick Start

Day 3-4: Quantization¶

• Study FP8 quantization on Hopper GPUs
• Learn INT8 and INT4 quantization techniques
• Understand AWQ and GPTQ post-training quantization
• Practice calibration and accuracy validation
• Reference: TensorRT-LLM Quantization

Day 5-7: Batching and Performance¶

• Study continuous batching and in-flight batching
• Learn KV cache management and paged attention
• Understand speculative decoding for latency reduction
• Practice benchmarking with TensorRT-LLM tools
• Lab: Compile and benchmark a model with different quantization levels

Week 5: RAG Pipelines¶

Day 1-2: RAG Architecture¶

• Study end-to-end RAG pipeline architecture
• Learn document ingestion - loading, parsing, chunking
• Understand embedding model selection and tradeoffs
• Explore vector database options (Milvus, FAISS, pgvector)
• Reference: NVIDIA RAG Pipeline Guide

Day 3-4: Retrieval Optimization¶

• Study chunking strategies - fixed-size, semantic, recursive
• Learn hybrid search - dense vector + sparse keyword
• Understand re-ranking with cross-encoder models
• Practice query expansion and reformulation
• Reference: NeMo Retriever

Day 5-7: RAG Evaluation and Production¶

• Study retrieval metrics - precision, recall, MRR, NDCG
• Learn generation evaluation - faithfulness, relevance
• Understand end-to-end RAG evaluation frameworks
• Practice building a complete RAG pipeline with NVIDIA tools
• Lab: Build and evaluate a RAG system end-to-end

Phase 3: Production and Exam Prep (Weeks 6-8)¶

Week 6: Production Deployment¶

Day 1-2: NVIDIA NIM Deployment¶

• Study NIM architecture and deployment options
• Deploy a model using NIM containers
• Learn NIM API endpoints (OpenAI-compatible)
• Understand NIM configuration and optimization
• Reference: NIM Getting Started

Day 3-4: Serving and Scaling¶

• Study load balancing for inference services
• Learn auto-scaling with Kubernetes HPA
• Understand Triton Inference Server for multi-model serving
• Practice A/B testing and canary deployments
• Reference: Triton Inference Server

Day 5-7: Safety and Monitoring¶

• Study NeMo Guardrails framework
• Learn input/output safety filtering
• Understand monitoring - latency, throughput, GPU metrics
• Practice implementing guardrails for an LLM application
• Reference: NeMo Guardrails

Week 7: Review and Practice Exams¶

Day 1-3: Domain Review¶

• Review all five domains systematically
• Create flashcards for key concepts and numbers
• Re-read NVIDIA documentation for weak areas
• Complete any unfinished labs

Day 4-5: Practice Questions¶

• Work through scenario-based practice questions
• Focus on questions combining multiple domains
• Review answers and identify remaining gaps
• Study common distractors and traps

Day 6-7: Gap Analysis¶

• Identify weakest domain areas
• Deep-dive into areas with lowest confidence
• Re-read relevant documentation sections
• Practice explaining concepts out loud

Week 8: Final Preparation¶

Day 1-3: Intensive Review¶

• Review fact sheet and key reference numbers
• Practice rapid-fire concept identification
• Review all NVIDIA tool configurations and commands
• Focus on production deployment patterns

Day 4-5: Final Practice¶

• Complete a full-length timed practice session (120 minutes)
• Review all incorrect answers
• Final review of weak areas

Day 6: Rest and Light Review¶

• Light review of fact sheet only
• Prepare exam logistics and setup
• Get good rest before exam day

Day 7: Exam Day¶

• Review key concepts briefly (30 minutes max)
• Verify exam environment and connectivity
• Take the exam with confidence
• Use flagging strategy for uncertain questions

Progress Tracking¶

Weekly Milestones¶

• Week 1-2: Understand transformer architecture, NeMo basics, distributed training
• Week 3: Master fine-tuning methods (LoRA, RLHF, data prep)
• Week 4: Proficient with TensorRT-LLM optimization and quantization
• Week 5: Build and evaluate RAG pipelines end-to-end
• Week 6: Deploy production systems with NIM, guardrails, monitoring
• Week 7: Complete practice questions, identify and fill gaps
• Week 8: Final review and exam

Self-Assessment Questions¶

• Can I explain the transformer architecture and its variants?
• Do I know when to use LoRA vs QLoRA vs full fine-tuning?
• Can I select the right quantization strategy for a given constraint?
• Can I design a production RAG pipeline with NVIDIA tools?
• Do I understand NIM deployment and scaling strategies?
• Can I implement guardrails for safe LLM applications?