NVIDIA Certified Professional – Generative AI LLMs (NCP-GENL)

Learning plan

About 80 hours of focused study, split into 5 phases. Pick your weekly pace and the timeline adjusts.

Estimated finish: 8 weeks

What the exam covers

Source: official exam guide. Weights can change, so check before you start.

Before you start

• NVIDIA recommends 2-3 years of practical AI/ML work with LLMs, covering transformers, distributed parallelism and parameter-efficient fine-tuning
• Strong Python. C++ helps for optimisation topics. Docker and Kubernetes basics are beneficial
• Access to at least one cloud GPU (an A10G/L4 class card is enough for the labs)

1. Weeks 1–2 · 12 h

   LLM architecture and prompt engineering

   • Explain decoder-only architecture, KV cache, attention variants (MHA, GQA) and how context length drives memory use
   • Apply zero/one/few-shot, chain-of-thought and structured-output prompting, and adapt a model to a new domain with prompting alone
   • Compare greedy, beam, temperature and top-p decoding and their effect on output control

   Use: Hugging Face LLM Course · NVIDIA DLI Building RAG Agents With LLMs

2. Weeks 2–3 · 16 h

   Data preparation and fine-tuning

   • Clean, deduplicate and filter a training corpus, and manage tokenisers and vocabulary for a domain
   • Choose between full fine-tuning, LoRA/QLoRA, P-tuning and SFT vs preference tuning (DPO/RLHF) by data size and budget
   • Fine-tune with NeMo or Hugging Face PEFT and avoid catastrophic forgetting

   Hands-on: QLoRA-fine-tune a 7-8B open model on a public instruction dataset on a single cloud GPU. Evaluate before and after on a held-out set and on a general benchmark slice to check for forgetting.

   Use: NVIDIA NeMo Framework user guide · Hugging Face PEFT docs · NVIDIA DLI Adding New Knowledge to LLMs (instructor-led)

3. Weeks 3–5 · 22 h

   Model optimization and GPU acceleration (31% combined)

   • Quantise models (FP8, INT8, INT4 AWQ/GPTQ) and measure the accuracy-vs-throughput trade-off
   • Build and serve an engine with TensorRT-LLM behind Triton, using in-flight batching and paged KV cache
   • Explain data, tensor, pipeline and sequence parallelism, and when each relieves a memory or compute bottleneck
   • Profile GPU kernels and memory with Nsight Systems/Compute and fix batch-size and memory bottlenecks

   Hands-on: Convert your fine-tuned model to a TensorRT-LLM engine in FP16 and INT4 AWQ, serve both with Triton, and record tokens/sec, time-to-first-token and an accuracy check at batch sizes 1, 8 and 32.

   Use: TensorRT-LLM on GitHub · Triton Inference Server user guide · Megatron-LM on GitHub (parallelism reference) · NVIDIA DLI Optimizing CUDA ML Codes With Nsight Profiling Tools

4. Weeks 6–7 · 18 h

   Deployment, evaluation, monitoring and safety

   • Deploy containerised inference on Kubernetes with autoscaling, and roll out new model versions safely
   • Design evaluation with quantitative metrics (perplexity, ROUGE, exact match), LLM-as-judge and human review, plus error analysis
   • Build dashboards for latency, throughput, error rates and drift, and set up retraining and versioning triggers
   • Audit outputs for bias, add guardrails, and monitor ongoing compliance

   Use: NVIDIA DLI Deploying RAG Pipelines for Production at Scale · NeMo Guardrails on GitHub

5. Weeks 7–8 · 12 h

   Practice and review

   • Re-quiz Model Optimization and GPU Acceleration first, since together they are almost a third of the exam
   • Practise 60-70 questions in 120 minutes and review every miss against the blueprint

   Use: NCP-GENL exam page and blueprint

You're ready to book when…

• You have fine-tuned, quantised and served a model with TensorRT-LLM and Triton yourself
• You can choose a parallelism strategy and a quantisation scheme for a given GPU budget and justify it
• You can read an Nsight trace and name the bottleneck
• You score 80% or higher on timed 60-70 question practice sets

Exam tips

• The exam is 60-70 questions in 120 minutes, online and remotely proctored. It is valid for two years and you renew by retaking it.
• The Fine-Tuning blueprint text on NVIDIA's page describes OpenUSD data interchange, which looks like a copy error. Study LLM fine-tuning as described in the page's Exam Preparation section.