This professional-level cheat sheet covers all 5 exam domains with optimization formulas, parallelism strategies, and deployment patterns. Based on official NVIDIA exam guide (26/22/22/18/12 weighting).
Test yourself before you memorise anything. These three questions come from the free NCP-GENL sampler and use the real exam format.
Three quick NCP-GENL questions
Question 1 of 3Model Optimization
Your latency-sensitive production LLM deployment uses TensorRT-LLM on Hopper GPUs and experiences high memory pressure with long context lengths. What is the recommended approach for KV cache optimization?
Pick one answer
AApply INT8 KV cache quantization to maximize the compression ratio per token
BEnable FP8 KV cache quantization to exploit the Hopper FP8 tensor cores
COffload the KV cache to host memory to free up GPU device memory
DTurn on KV cache block reuse so shared prompt prefixes are not recomputed
Answer: B
On Hopper (and Ada) GPUs, FP8 KV cache quantization is the recommended path: it shrinks the cache in place for long contexts while keeping accuracy impact lower than INT8 in most tested cases, and it maps directly onto Hopper's native FP8 support without adding data-movement latency.
Question 2 of 3Data Preparation
What is the primary advantage of using SentencePiece for tokenization compared to word-level tokenization?
Pick one answer
AIt treats whitespace as a reserved symbol so detokenization is reversible across any language
BIt applies the unigram language-model objective to prune the most frequent whole words first
CIt handles many languages and rare words through language-agnostic subword segmentation
DIt guarantees a fixed token count per input regardless of the configured vocabulary size
Answer: C
SentencePiece implements subword algorithms (BPE and unigram LM) directly on Unicode characters without language-specific pre-tokenization or whitespace assumptions. By splitting words into subword units it gracefully covers out-of-vocabulary and rare terms, which is exactly what word-level tokenization cannot do, and it works across diverse languages and scripts.
Question 3 of 3Prompt Engineering & Advanced Techniques
Your team is building a customer service chatbot that needs to follow a specific troubleshooting workflow. The LLM sometimes skips steps or hallucinates solutions. Which prompt engineering technique would most effectively ensure the model follows the required step-by-step reasoning process?
Pick one answer
AAdd a retrieval-augmented generation layer so each answer is grounded in the troubleshooting knowledge base of articles
BApply self-consistency decoding by sampling several independent reasoning chains and majority-voting the result
CSwitch to greedy decoding at low temperature so the model returns one deterministic, repeatable response
DUse few-shot Chain-of-Thought prompts that demonstrate the full troubleshooting workflow as worked examples
Answer: D
Few-shot CoT supplies concrete worked examples that walk through every troubleshooting stage in order, so the model learns the exact step-by-step pattern to imitate. Demonstrating the complete reasoning process directly targets both failure modes: it enforces adherence to the required workflow and reduces hallucinated solutions by anchoring each step to the shown pattern.
Full NCP-GENL set: 7 timed exams, every answer explained, $19.99 one-time.
ZeRO-2: Memory constrained, acceptable comm overhead
ZeRO-3: Extreme memory needs, willing to trade speed
GPU Memory Calculation
Training Memory Estimate
Memory = Model + Optimizer + Gradients + Activations
Copy
Gradient Checkpointing
Trade-off: Memory ↔ Compute
Without checkpointing: Store all activations → High memory
With checkpointing: Store checkpoints, recompute between → 30-50% memory reduction, 20-30% compute increase