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

This part covers how an LLM is built and how you adapt it to new tasks through prompts.

• Encoders, decoders and self-attention in transformer models
• Embeddings and sampling in text generation
• Chain-of-thought, zero-, one- and few-shot prompts
• Output control

Here you prepare data and adapt a pretrained model to a domain.

• Cleaning, curating and formatting datasets
• Tokenisation and vocabulary
• Fine-tuning and domain adaptation
• Parameter-efficient methods such as LoRA

The largest block of the exam: models get smaller and faster, and training and inference scale across several GPUs.

• Pruning, quantisation and knowledge distillation
• Multi-GPU and distributed training with data, tensor and pipeline parallelism
• Memory and batch optimisation
• Finding bottlenecks through profiling

This part takes an LLM into production and keeps it reliable there.

• Containerised inference pipelines with NVIDIA Dynamo-Triton, formerly Triton Inference Server
• Model serving and orchestration with Kubernetes
• Monitoring dashboards, reliability metrics and root-cause analysis
• Versioning and retraining in operation

Finally you measure the quality of a model and limit its risks.

• Benchmarks, metrics and error analysis
• Scalable evaluation
• Checking bias and fairness
• Guardrails and compliance
• Monitoring for ethical compliance in operation