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

What the NVIDIA-Certified Professional exam covers

• LLM Architecture45 questions
• Prompt Engineering98 questions
• Data Preparation68 questions
• Model Optimization128 questions
• Fine-Tuning97 questions
• Evaluation53 questions
• GPU Acceleration and Optimization105 questions
• Model Deployment67 questions
• Production Monitoring and Reliability52 questions
• Safety, Ethics, and Compliance38 questions

Free NVIDIA-Certified Professional sample questions

A sample of 10 questions with answers and explanations. Sign up free to practice all 751.

1. Question 1LLM Architecture

   Which pair of sublayers makes up one decoder transformer block?

   • ASelf-attention followed by a position-wise feed-forward networkCorrect
   • BCross-attention followed by a recurrent gating unit over the sequence
   • CA convolution over the sequence followed by a pooling operation
   • DAn embedding lookup followed by a normalisation of the vocabulary
   ✓ Correct answer: A

   The block mixes information across positions in the attention sublayer and then transforms each position independently in the feed-forward sublayer, and every decoder-only architecture repeats that pair. Cross-attention with a recurrent gating unit, a convolution followed by pooling and an embedding lookup followed by vocabulary normalisation describe an encoder-decoder stage, a convolutional network and a preprocessing step rather than a decoder block.

   Why the other options are wrong

   • BCross-attention appears in an encoder-decoder stage, not a decoder-only block.
   • CA convolution over the sequence belongs to a different family of architecture.
   • DAn embedding lookup happens once before the stack rather than inside a block.
2. Question 2Prompt Engineering

   Which condition does that technique require to work?

   • AA final answer that can be compared for equality across attemptsCorrect
   • BA temperature of zero, so every attempt follows the same path
   • CA single attempt, since additional attempts would disagree
   • DA model small enough that several attempts fit in one batch
   ✓ Correct answer: A

   Voting needs a value that two attempts can be said to agree on, which is why it suits arithmetic and classification far better than open prose. Zero temperature would make every attempt identical, a single attempt has nothing to vote with, and model size affects the cost rather than the validity.

   Why the other options are wrong

   • BZero temperature produces identical attempts with nothing to compare.
   • CA single attempt provides no agreement signal at all.
   • DModel size affects the cost of the attempts rather than the method.
3. Question 3Data Preparation

   Which describes benchmark contamination in a training corpus?

   • AEvaluation items appear in the training data, so the reported score is inflatedCorrect
   • BEvaluation items are missing from the training data, so the score is deflated
   • CEvaluation items are formatted differently from the training data
   • DEvaluation items outnumber the training items in a particular domain
   ✓ Correct answer: A

   A model that has seen the answers reports a capability it does not have, which is the most consequential data defect in the field. Items missing from training is the normal state, a formatting difference is a preparation matter, and a count imbalance describes a mixture problem.

   Why the other options are wrong

   • BEvaluation items absent from training is exactly the intended state.
   • CEvaluation items formatted differently is a preparation matter, not leakage.
   • DEvaluation items outnumbering training items describes a mixture problem.
4. Question 4Model Optimization

   Which measurement establishes whether a quantised model is acceptable?

   • ATask performance on the real workload, compared against the unquantised modelCorrect
   • BThe size of the model file, compared against the unquantised model
   • CThe average difference between quantised and original weight values
   • DThe number of layers that were successfully converted without error
   ✓ Correct answer: A

   Weight-level error correlates poorly with what a user experiences, so the only figure worth acting on is the one measured on the work the model actually does. File size, average weight error and a count of converted layers all confirm that the conversion happened rather than that it was harmless.

   Why the other options are wrong

   • BThe file size confirms the conversion happened rather than its effect.
   • CAverage weight error correlates poorly with task performance.
   • DA count of converted layers says nothing about the resulting quality.
5. Question 5Model OptimizationSelect all that apply

   Which TWO does continuous batching improve on that?

   • AA finished sequence's slot is taken by a waiting request immediatelyCorrect
   • BA newly arrived request need not wait for the current batch to endCorrect
   • CThe model produces higher quality output for each request served
   • DThe memory each sequence's cache occupies is reduced
   ✓ Correct answer: A, B

   Both come from treating batch membership as something that changes between steps rather than being fixed at formation. Output quality is unaffected by scheduling, and the cache each sequence occupies is a function of its length rather than of how the batch is managed.

   Why the other options are wrong

   • CScheduling does not change the quality of what the model produces.
   • DEach sequence's cache depends on its length rather than the batching.
6. Question 6Fine-TuningSelect all that apply

   Which TWO advantages does that simplification bring?

   • AFewer moving parts to tune, since there is no scorer to train separatelyCorrect
   • BA more stable run, since no reinforcement learning loop has to convergeCorrect
   • CA larger behavioural change than the classical pipeline can achieve
   • DNo need for preference data, since the objective is derived analytically
   ✓ Correct answer: A, B

   Both of these come from collapsing a fragile two-stage procedure into a single supervised-style objective, which is why the method has been so widely adopted. A larger behavioural change than the classical pipeline achieves is not something the method claims, and preference data remains exactly the input the objective is computed over.

   Why the other options are wrong

   • CThe size of the change is not what the simplification improves.
   • DPreference data is exactly what the objective consumes.
7. Question 7Evaluation

   Which removes position bias from a pairwise evaluation?

   • AJudging each pair twice with the order swapped and combining the verdictsCorrect
   • BJudging each pair once with the stronger system always shown first
   • CJudging each pair once with the answers in alphabetical order
   • DJudging each pair once and noting the order in the results
   ✓ Correct answer: A

   Running both orderings and combining them cancels the effect rather than merely documenting it, at twice the judging cost. Always showing the stronger first builds the bias in, alphabetical ordering is arbitrary but still fixed, and noting the order records the problem without correcting it.

   Why the other options are wrong

   • BAlways showing the stronger system first builds the bias into the result.
   • CAlphabetical ordering is arbitrary and still a fixed order.
   • DNoting the order documents the bias without removing it.
8. Question 8GPU Acceleration and Optimization

   Which deployment does tensor parallelism suit?

   • ADevices inside one machine connected by a high-bandwidth local linkCorrect
   • BDevices in separate machines connected by ordinary network hardware
   • CDevices in separate data centres connected over a wide area link
   • DA single device, since the split is logical rather than physical
   ✓ Correct answer: A

   The frequency of the exchange means only the fastest available link keeps the overhead tolerable, which in practice means devices sharing a chassis. Separate machines, separate data centres and a single device describe links far too slow for the traffic and a case where no split is needed.

   Why the other options are wrong

   • BOrdinary network hardware is far too slow for a per-layer exchange.
   • CA wide area link is orders of magnitude too slow for this pattern.
   • DA single device needs no split at all, whether logical or physical.
9. Question 9Model Deployment

   Which describes the weights relative to the container image?

   • AUsually kept outside it, because an image of that size is unwieldy to moveCorrect
   • BUsually kept inside it, because that guarantees they match the code
   • CAlways kept inside it, because a container cannot read external storage
   • DAlways kept outside it, because images cannot exceed a fixed size
   ✓ Correct answer: A

   Baking tens of gigabytes into an image makes every pull enormous and every code change a full re-upload, so the weights are usually fetched at startup instead. Guaranteeing a match by inclusion is a real benefit that rarely outweighs the size, containers read external storage routinely, and no fixed size limit forces the decision.

   Why the other options are wrong

   • BInclusion does guarantee a match, at a size that is rarely worth it.
   • CA container reads external storage routinely.
   • DNo fixed image size limit forces the weights outside.
10. Question 10Production Monitoring and Reliability

    Which symptom-level alert suits this service best?

    • AThe share of requests exceeding the latency target, over a rolling windowCorrect
    • BThe device utilisation on any individual replica
    • CThe number of replicas currently registered with the load balancer
    • DThe size of the model artefact currently deployed
    ✓ Correct answer: A

    A rising share of requests missing their target is what a user experiences, whatever caused it, and it fires for causes nobody listed in advance. Device utilisation, a replica count and the artefact size describe internal state that can look fine while users suffer.

    Why the other options are wrong

    • BDevice utilisation can look healthy while users are queueing.
    • CThe number of replicas registered says nothing about whether users are served.
    • DThe artefact size is static and unrelated to the experience.

Who this NVIDIA-Certified Professional practice exam is for

This practice set is for anyone preparing for the NVIDIA-Certified Professional: Generative AI LLMs (NCP-GENL) exam at the advanced level - from first-time candidates building a foundation to experienced NVIDIA practitioners doing a final review before test day. If you learn best by working through realistic questions and reading why each answer is right or wrong, it is built for you.

How to use this NVIDIA-Certified Professional practice exam

1. Start with the free sample questions above to gauge your current baseline.
2. Read the full explanation on every question, including why each wrong option is wrong.
3. Track your weak domains and focus your study where you are losing the most marks.
4. Once you are scoring consistently well, take a timed, full-length mock exam.
5. Use your readiness score to decide when you are ready to book the real NVIDIA-Certified Professional exam.

Related NVIDIA resources

• NVIDIA-Certified Professional study guideKey concepts
• NVIDIA practice examsAll NVIDIA
• Certification pathWhere this fits
• Certification exam guides & tipsBlog
• Plans & pricingFree & paid
• NVIDIA-Certified Associate practice examRelated
• NVIDIA-Certified Associate practice examRelated
• NVIDIA-Certified Associate practice examRelated

NVIDIA-Certified Professional practice exam FAQ