Loading demo links...
What to study for each NCP-GENL domain
LLM Architecture
Covers transformer architecture in depth, including self-attention, positional encoding approaches such as RoPE, and decoder-only model design. It also addresses architectural variants and mechanisms such as mixture-of-experts and grouped-query attention that shape capability, context handling, and inference cost.
Key skills
Self-attention mechanics, RoPE and other positional encodings, mixture-of-experts, context-window design, and architecture choices across modern LLM families.
Prompt Engineering
Covers designing effective prompts for production LLM systems, including instruction design, few-shot examples, and chain-of-thought reasoning. It also addresses structured output, retrieval-grounded prompting, sampling controls, context management, and defending prompts against injection and misuse.
Key skills
Instruction design, few-shot example selection, chain-of-thought and structured output control, prompt templating, and defending against prompt injection.
Data Preparation
Covers curating and preparing data for LLM training, fine-tuning, and retrieval, including collection, cleaning, deduplication, and quality filtering. It also addresses tokenisation, personally identifiable information handling, benchmark decontamination, and chunking strategies for retrieval corpora.
Key skills
Collecting and cleaning pretraining and fine-tuning corpora, tokenization and vocabulary management, deduplication, and dataset quality filtering.
Model Optimization
Covers making LLM inference and training faster and cheaper, including quantisation methods such as GPTQ and AWQ, pruning, and knowledge distillation. It also addresses memory and compute optimisations such as FlashAttention, KV cache optimisation, speculative decoding, activation checkpointing, and sharded training.
Key skills
Quantization (GPTQ, AWQ, FP8), pruning and distillation, KV-cache optimisation, speculative decoding, and reducing inference cost without losing quality.
Fine-Tuning
Covers adapting pretrained LLMs to specific tasks and behaviours through supervised fine-tuning on instruction and response data. It also addresses parameter-efficient methods such as LoRA and QLoRA, preference alignment with RLHF and DPO, hyperparameter choices, and preventing catastrophic forgetting and overfitting.
Key skills
Supervised fine-tuning, parameter-efficient methods (LoRA/QLoRA), RLHF and DPO alignment, catastrophic forgetting, and fine-tuning hyperparameters and data mixing.
Evaluation
Covers measuring LLM quality with task metrics, benchmark suites, human evaluation, and LLM-as-judge methods, including their biases and limitations. It also addresses retrieval and groundedness evaluation for RAG systems, benchmark contamination, and maintaining golden regression sets.
Key skills
Task metrics and benchmark suites, human evaluation design, LLM-as-judge methods and their biases, regression testing, and evaluating for safety.
GPU Acceleration and Optimization
Covers accelerating LLM workloads on NVIDIA GPUs, including Tensor Core precision formats, CUDA Graphs, and kernel and operator optimisation. It also addresses parallelism strategies such as tensor, pipeline, and data parallelism, and serving optimisations in TensorRT-LLM and vLLM including PagedAttention and continuous batching.
Key skills
Tensor Core precision formats, kernel fusion and CUDA Graphs, tensor/pipeline parallelism, profiling with NVIDIA tools, and memory and batch optimisation.
Model Deployment
Covers deploying LLMs into production serving environments, including model registries and versioning, inference servers such as Triton, and streaming responses. It also addresses rollout strategies such as canary and shadow releases, autoscaling on GPU-aware signals, and latency metrics such as time to first token.
Key skills
Inference serving with TensorRT-LLM and vLLM-style engines, model registries and versioning, autoscaling, and safe rollout strategies.
Production Monitoring and Reliability
Covers operating LLM systems reliably in production, including monitoring quality, latency tails, token usage, and cost, and detecting input and output drift. It also addresses safe logging with redaction, feedback signals, fallbacks, circuit breakers, and incident response for model and provider failures.
Key skills
Monitoring output quality and latency tails, token-usage and cost tracking, fallback and retry design, and incident response for LLM services.
Safety, Ethics, and Compliance
Covers responsible deployment of LLM systems, including input and output guardrails, red teaming for jailbreaks and harmful content, and bias evaluation across user groups. It also addresses training-data provenance, licensing and copyright considerations, and privacy and compliance obligations.
Key skills
Input/output guardrails, red teaming for jailbreaks and data extraction, PII handling, bias evaluation, and regulatory compliance for deployed LLMs.