NCP-GENL
Free NVIDIA-Certified Professional: Generative AI and LLMs Practice Test
12 exam-style questions with full explanations — no sign-up. Score yourself, then close your gaps with the full course.
Exam fee ~$2006 exam domainsLevel Advanced2 timed practice tests in the course
✅ Free practice — no sign-up📝 Real exam-style questions💡 Detailed explanations💸 30-day money-back via Udemy
A team is deploying a 70B-parameter LLM for real-time chat inference on GPUs with limited KV-cache memory bandwidth. They need to substantially reduce memory traffic during autoregressive decoding while keeping output quality close to standard multi-head attention. Which attention variant BEST meets this requirement?
GQA groups query heads into clusters that share KV projections, cutting KV-cache memory and bandwidth substantially versus full MHA while retaining most heads' representational diversity, giving a much smaller quality drop than MQA's single shared KV pair.
A financial news aggregator answers customer questions about market events that change every hour. The knowledge base is updated continuously, answers must cite the specific source article, and the team cannot afford to retrain a model every hour. Which adaptation strategy BEST meets these requirements?
RAG decouples knowledge freshness from model weights: updating the vector index is near-instant and cheap compared to retraining, and retrieved passages naturally provide citable sources for grounding the answer.
A team needs to adapt a 70B-parameter foundation model to a specialized domain using only 500 labeled examples, and their entire training budget is a single 80GB A100 GPU. Which customization approach BEST meets these constraints?
QLoRA quantizes the frozen base model weights to 4-bit precision and trains only small LoRA adapter matrices in higher precision, dramatically reducing memory footprint so a 70B model can be fine-tuned on a single 80GB GPU, and it works well with small labeled datasets.
A team is training a 70B-parameter dense model across 4 nodes with 8 GPUs each (32 GPUs total, NVLink within node, InfiniBand between nodes). The model does not fit on a single GPU even with sharded optimizer states, and the team wants to minimize expensive inter-node communication. Which parallelism configuration BEST fits this hardware topology?
Tensor parallelism requires frequent, low-latency all-reduce communication, so it should be confined to the high-bandwidth NVLink domain within a node. Pipeline parallelism only exchanges activations at stage boundaries, making it tolerant of the lower-bandwidth inter-node InfiniBand links. This 3D-parallelism pattern (TP intra-node, PP inter-node, DP for scaling) is the standard NVIDIA-recommended topology-aware strategy.
A team serves an LLM with NVIDIA Triton Inference Server behind a Kubernetes-based NIM deployment. They need to roll out a new fine-tuned model version to production while minimizing risk of regressions, and they want the ability to automatically roll back based on live error-rate metrics without manual intervention. Which deployment strategy BEST meets these requirements?
Canary rollouts combined with automated metric-based rollback provide progressive exposure and closed-loop safety without requiring a human to watch dashboards, directly satisfying the 'automatic rollback based on live metrics' requirement.
A team fine-tuned an LLM for a medical Q&A chatbot. During testing, the model confidently generates plausible-sounding drug dosages that are not supported by any retrieved source document. Which evaluation approach BEST detects this specific failure mode before production deployment?
Hallucination detection specifically evaluates whether generated claims are grounded in and supported by retrieved/source context, which is exactly the failure described (confident but unsupported factual claims).
An organization wants to increase total model capacity to handle multiple domains (code, math, general chat) but must keep per-token inference compute close to that of a much smaller dense model. Which architecture choice BEST satisfies this constraint?
MoE architectures activate only a small subset of experts per token via a learned router, so total parameter count (capacity) can scale into the hundreds of billions while active FLOPs per token stay near a much smaller dense model's cost.
A legal document classification service must apply a very specific, consistent tagging taxonomy to every document, has 50,000 labeled examples available, and requires sub-100ms inference latency in production with no external calls. Which approach is MOST appropriate?
With abundant labeled data, a fixed consistent output format, and a hard low-latency/no-external-call constraint, fine-tuning bakes the behavior directly into the weights, eliminating retrieval latency and long reasoning chains while maximizing consistency.
A company wants to align a model's outputs to human preference data but does not have the resources to train a separate reward model and run PPO-based reinforcement learning. Which preference optimization technique should they use instead?
DPO directly optimizes the policy model on preference pairs (chosen vs. rejected responses) using a closed-form loss derived from the RLHF objective, eliminating the need for a separate reward model or an RL training loop like PPO.
Which post-training weight quantization technique identifies salient weight channels based on activation magnitude statistics and protects those channels from precision loss, rather than minimizing layer-wise weight reconstruction error via a Hessian-based approach?
AWQ (Activation-aware Weight Quantization) observes activation magnitudes to determine which weight channels are most salient and scales/protects those channels, avoiding the costly per-layer Hessian-based reconstruction optimization that GPTQ performs.
An LLM inference service on Triton Inference Server receives many concurrent small requests. GPU utilization is low despite high request volume, and per-request latency is acceptable but overall throughput is far below the GPU's theoretical capacity. Which Triton feature should be enabled FIRST to increase throughput without changing the model itself?
Dynamic batching aggregates concurrent requests into larger GPU batches, improving GPU utilization and throughput, which is exactly the symptom described (low utilization despite high request volume).
An organization needs to evaluate the quality of open-ended, creative summaries generated by an LLM at scale, where no single reference answer exists and nuanced judgment (coherence, relevance, style) is required. Which evaluation method is MOST appropriate for this requirement?
LLM-as-judge is designed for open-ended generation tasks lacking a single correct answer, using a rubric to approximate nuanced human-like judgment at scale, which fits creative summarization evaluation.
Ready for the real thing?
The full course: two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed explanations.
$129.99 $34.99 with code FREETEST33 — valid through September 7.
Get my $34.99 deal →NCP-GENL exam — quick answers
How much does the NCP-GENL exam cost?
The exam fee is approximately $200 and varies by region — confirm current pricing with the certification vendor before you book.
What topics are on the exam?
It covers 6 domains: LLM Foundations and Architecture (~15%), Prompt Engineering and Adaptation (~15%), Data Preparation and Fine-Tuning (~18%), Optimization and Acceleration (~20%), Deployment, Serving, and Monitoring (~18%), Evaluation and Responsible AI (~14%). The full course has a dedicated chapter, lab and practice-test coverage for each.
Is this practice test really free?
Yes — all questions on this page are free with explanations and no sign-up. The paid Udemy course adds two full-length timed exams, video lessons and hands-on labs.
How do I get the discount?
Use code FREETEST33 at checkout for $34.99 (list $129.99) through September 7 — the enroll button applies it automatically.
Will this prepare me for the real exam?
The questions mirror the real exam's style and are mapped to the official domains. This is exam-focused preparation — combine the free test with the full course's timed simulations to gauge your readiness.