NCA-GENL - Prep Course - GenAI & LLMs Associate

Discover why companies invest in generative AI and large language models to cut time and costs, boost decision making, and drive innovation, advancing your career.

Explore Nvidia certification tracks for infrastructure professionals and developers. Learn Gen L for generative AI and llms across associate and professional levels.

Explore the NCA-GENL certification with associate versus professional levels, covering generative AI and large language models concepts, exam logistics, and foundational topics like model training, customization, and rag prompt engineering.

Explore AI infrastructure, AI/ML fundamentals, generative AI and LLMs, transformer architecture, model selection and customization, training, and the Nvidia ecosystem, all while covering ethical AI and exam prep.

Explore the AI ML stack from infrastructure to applications, focusing on model selection, data preprocessing, and MLOps orchestration to customize and fine-tune models for on prem or cloud use cases.

Explore how graphics processing units power parallel computations for machine learning, from early GPU origins and CUDA to the rise of AlexNet and GPU-accelerated computing.

Explore the CPU vs GPU difference through an analogy and compare cores, parallel processing, latency, and throughput. Learn GPU use cases in graphics rendering, AI simulation, and mining.

Compare CPU and GPU architectures, noting CPUs use a few cores with private and shared caches, while GPUs hold thousands of cores and large memory for parallel processing.

Memory pooling preallocates and reuses memory to keep CPU-GPU data transfers efficient, reducing bottlenecks between system memory and GPU RAM.

Compare cloud and on-prem infrastructure for AI workloads, weighing cloud's low entry cost and scalability against on-prem's data security, sovereignty, and upfront investment.

Discover how AI data centers use high-speed network infrastructure, comparing InfiniBand and Ethernet for scalable, use-case driven inter-machine communication.

Explore storage infrastructure for AI workloads, balancing high throughput and low latency with nvme ssd, parallel file systems across multiple gpu nodes, and object storage in a tiered data center.

Identify AI infrastructure components, including CPU and GPU compute, networking options like Ethernet or InfiniBand, and storage choices, and compare cloud versus on-prem setups.

Explore AI and ML fundamentals, align key terms, and show how machine learning enables pattern recognition, inference, and decision making and how these ideas appear in daily AI applications.

Explain deep learning as learning from data patterns with artificial neural networks inspired by the human brain, and illustrate how layered networks transform inputs into predictions.

Contrast rule-based recommendations with machine learning by using purchase history, age, and location to generate personalized product suggestions for online stores.

Explore how a trained algorithm identifies patterns in data without explicit rules, and learn how training data turns it into a model that predicts and recommends.

Learn how to train a model through a coffee-shop analogy, mapping data to ingredients, a recipe to the algorithm, GPUs to hardware, and inference to production predictions.

Review foundational AI concepts, including artificial intelligence, machine learning, deep learning, and neural networks, and explain models, algorithms, and inference on unseen data.

Explore foundation models as general-purpose AI trained on vast, diverse, and unlabeled data, adaptable to multiple tasks and modalities, and usable as a base for downstream applications.

A large language model processes vast text to generate human-like text and perform tasks such as answering questions, summarizing, and translating, powered by high-quality data and a robust transformer architecture.

Trace the road to generative AI by exploring word embedding, vector space, and the evolution from word2vec to language models that power NLP tasks like text summarization and sentiment analysis.

Understand how transformers use attention to model context for next-word prediction, and see how this mechanism enables generative AI to produce sentences, paragraphs, and even new models.

Explore how generative AI uses foundational models trained on unlabeled data, and how large language models generate human-like text, with word embedding like word2vec and transformer selection customization.

Unfold the power of transformers, the engine behind modern AI, by turning long sequences into other sequences with self-attention.

Discover how a transformer completes a sentence through five phases—tokenization, encoding, word embedding, decoding with self-attention, and output generation—using puppy is to dog as kitten is to cat.

Explore how tokenization breaks text into tokens and subword bricks, enabling efficient encoding, decoding, and faster processing with byte pair encoding, SentencePiece, and WordPiece.

The lecture shows tokenization converts words into numbers via token IDs in the model's vocabulary, and explains that embedding and vectorization are needed to reveal semantic meaning.

Convert tokens after tokenization into dense vector embeddings with positional encoding, then apply self-attention in a transformer to map words into a contextual vector space that captures meaning.

Explore how decoding predicts the next word using attention and learned patterns, converts it to a vector, and generates output for sentences and paragraphs.

See how models convert tokens back into text, with subword token encoding, word embedding, and decoding to generate output. Preview transformer concepts such as positional encoding, encoder–decoder architecture, and self-attention.

Explore positional encoding and how a unique numbered sticker on each token preserves word order, helping the model understand who chased whom and avoid a bag-of-words view.

Explore how transformers generate the next word using tokenization, embedding, and multi-head attention. Visualize the transformer explainer, view embedding, attention, and softmax probabilities behind models like GPT two model.

Explore how the attention mechanism powers transformers and self-attention, using group chat analogies to show how positional encoding and highlighted context enable accurate text generation.

Transformers use multi-head attention, with several specialist heads highlighting different aspects of the same chat, then combine outputs to map relationships and enable parallel processing with GPUs.

Explore the transformer’s encoder and decoder blocks, their distinct use cases for search and generation, and how encoder-decoder models translate and summarize.

Choose encoder-only for understanding tasks like sentiment analysis and spam detection; decoder-only for generation tasks such as chat, emails, and code; encoder-decoder blends handle translation and summarization.

Learn how transformer architecture underpins modern language models and multi-stage training, then compare auto regressive, bidirectional, denoising, and sequence-to-sequence prediction strategies.

Explore transformer architecture, including attention mechanisms, positional encoding, and multi-head attention, to understand how neural networks model sequences in NVIDIA GenAI & LLMs.

Define your need and use case to narrow model options, shortlist candidates, compare model cards for capabilities, costs, risks, benchmarks, and decide between small and large language models.

Compare small language models and large language models, noting shared unlabeled pre-training and transformer architectures. Differentiate latency, memory, and throughput, and show how veracity favors large models.

Assess models by accuracy, reliability, and performance to select the best option for your use case. Compare bleu score, rouge score, and cosine similarity for translations and summaries.

Explore rouge as a recall-based metric for content coverage. Leverage cosine similarity to measure semantic meaning between texts.

Compare blue, rouge, and cosine similarity as key metrics for machine translation accuracy, summarization evaluation, and semantic search, highlighting their use cases and limitations.

Learn to test and compare AI models using cross validation, side-by-side output comparisons, AB testing, and pilots, with fold-based evaluation to gauge generalization.

Compare explore and exploit in a/b testing by randomly assigning users to old and new models, balancing risk and learning while controlling for external factors to support data-driven optimization.

Select the best model for your use case by balancing performance, cost, ease of use, and reliability. Consider affordability, deployability, risk, and documentation, then re-evaluate choices over time with testing.

Compare models by selecting SLM versus LM, assess accuracy with rouge metrics and cosine similarity, and use cross-validation and a/b testing to optimize GenAI performance.

Explore why customizing a model matters, from domain language in healthcare and finance to better context awareness for customer service, and compare prompt engineering, RAG, and transfer learning.

Learn prompt engineering to customize model output, using zero-shot, one-shot, and few-shot prompting, plus chain-of-thought reasoning, with practical real-world examples.

Learn how system prompts drive model behavior with role-based instructions and personas, enforce rules and formats, and apply zero-shot, one-shot, few-shot, and chain-of-thought prompting techniques.

Explore retrieval augmented generation (rag) and how to augment llm outputs with a vector database and knowledge base to provide up-to-date, sourced responses and reduce hallucination.

Explore retrieval augmented generation (rag) by combining user queries with external knowledge sources, chunking data into chunks, retrieving them, and augmenting prompts for the foundation model to generate accurate responses.

Compare prompt engineering, rag, agents, and transfer learning for varied tasks. Apply a decision tree to guide when to use each approach for up-to-date data, static info, or industry knowledge.

Explore transfer learning from pre-trained models, reusing features and adapting to new tasks through fine-tuning, knowledge distillation, and deployment optimization with pruning and quantization.

Transfer learning reuses a pre-trained model to adapt to a new related task, saving data and training time while improving accuracy in vision, NLP, speech, and autonomous driving.

Fine-tuning retrains a pre-trained model on task-specific data to improve performance, updating the final layer or top layers for a new task, with examples like emails and x-ray fracture detection.

Learn how knowledge distillation transfers knowledge from a large teacher model to a smaller student model, enabling faster, lighter inference with similar accuracy for edge devices like mobile face recognition.

Split your data into training and validation sets, typically 80/20, and measure validation accuracy on unseen data to ensure generalization after fine-tuning or distillation.

Explore techniques for customizing foundation models, including prompt engineering with zero/one/few-shot prompting, chain of thought, rag, transfer learning, fine tuning, and knowledge distillation.

Navigate the five steps in gen AI and ML projects—data collection, cleaning, and preparation, model training, evaluation, and deployment via API or inference server—highlighting data quality.

Collect data from public data sets, enterprise data, and external APIs to support model development. Ensure ethical, privacy-aware handling with compliant storage and access controls.

Learn to use regular expressions to filter and extract specific data from logs, such as errors and warnings, for fast pre-cleaning before heavier ML processing, with noted pros and cons.

Perform eda (exploratory data analysis) to understand data structure, quality, and patterns before modeling. Use histograms, scatter plots, and correlation matrices to reveal outliers and guide feature selection and pre-processing.

Explore essential EDA techniques, focusing on n-gram summaries and word frequency analysis, and use descriptive statistics and visualizations like histograms and box plots to reveal patterns and sentiment.

Explore how model training turns pre-processed data into a functional intelligent model by adjusting internal parameters to minimize error, covering model selection, training strategy, hyperparameters, execution, and monitoring.

Choose the right base model based on the problem, data, modality, and constraints. Apply training strategies from scratch or fine-tuning, and monitor hyperparameters and progress to optimize performance.

Evaluate a trained model on unseen data to decide deployment readiness, serving as a decision checkpoint for data improvements and retraining, using BLEU, ROUGE, accuracy, and qualitative reviews.

Deploy the trained model to production by integrating it with apps via API for real-time or batch use, ensuring low latency, high throughput, fault tolerance, and handling unseen data.

Explore deployment options for genai and llms, including batch offline predictions and real-time processing, plus cloud vs on-prem hosting, containers, docker, and ci/cd pipelines.

Convert your model to ONNX to make it framework agnostic and portable across deployment targets, from PyTorch to TensorRT or edge devices.

Learn how quantization reduces precision from 32-bit to 8-bit or 16-bit, shrinking model size and speeding inferences with minimal accuracy loss, lowering power for mobile and edge devices.

Explore the model training lifecycle from data collection and processing to evaluation and deployment, including hyperparameters, onnx portability, and quantization for efficient inference.