Free NCP-GENL: NVIDIA Certified Professional — Generative AI LLMs Exam Questions and Answers

Which description matches one-shot prompting?

• AA. The prompt includes a single worked example of the task before the actual input
• BB. The prompt gives the task instruction and no demonstrations of the desired output
• CC. The model weights are updated using a single labeled example before inference
• DD. The model is allowed to generate only one output token for each request

An engineer implements a small decoder-only transformer from scratch. During training the loss drops to nearly zero within a few hundred steps, but autoregressive generation produces incoherent text. Reviewing the attention code, the engineer finds that the score matrix is passed to softmax without any masking. What is the cause of the behavior?

• AA. Each position can attend to later tokens, so the model sees the target it is supposed to predict
• BB. Softmax was computed across attention heads instead of across keys, which makes the weights uniform
• CC. Positional encodings were added to the values instead of the inputs, so token order is lost
• DD. Layer normalization is applied after the residual connection, which scales attention logits too low

Which NVIDIA toolkit lets developers add programmable input, output and dialog rails to an LLM-based application?

• AA. Triton Inference Server
• BB. TensorRT-LLM
• CC. NeMo Guardrails
• DD. NeMo Curator

Which metric estimates the probability that at least one of k sampled programs passes a problem's unit tests when evaluating a code generation model?

• AA. ROUGE-L
• BB. BLEU
• CC. pass@k
• DD. Perplexity

A team compares two chat models with an LLM judge. The judge prompt always lists Response 1 followed by Response 2. Model A wins 68% of pairs when its answer is shown as Response 1, but only 41% when its answer is shown as Response 2. The responses from both models have similar lengths. Which diagnosis and fix is most appropriate?

• AA. Sampling noise; rerun the judge at temperature 1.0 and average the win rate over five runs
• BB. Self-preference bias; swap the judge for a smaller model fine-tuned on the candidate outputs
• CC. Verbosity bias; cap every response at the same token count before the judge scores the pair
• DD. Position bias; judge each pair in both orders and count a win only when both verdicts agree

Which statement best describes NVIDIA NIM?

• AA. A set of prebuilt containers that package a model with an optimized inference engine behind a standard API
• BB. A Kubernetes operator that schedules GPU resources and replaces the NVIDIA device plugin
• CC. A training framework that shards model weights across nodes using tensor and pipeline parallelism
• DD. A post-training toolkit that quantizes model checkpoints to FP8 or INT4 formats

A decoder-only model uses sliding window attention with a window of 4,096 tokens in each of its 48 layers. Approximately how many tokens back can information from an input token influence the final-layer representation of the current token?

• AA. 4,096
• BB. 49,152
• CC. 98,304
• DD. 196,608

An LLM is served by NVIDIA Triton Inference Server pods in a Kubernetes Deployment. Loading a new model version takes about four minutes per pod. The team must publish the update with no failed client requests. Which configuration BEST supports this?

• AA. Use a rolling update with maxUnavailable 0 and a readiness probe on Triton's /v2/health/ready endpoint
• BB. Use a liveness probe on the Triton port and restart pods when they fail to respond
• CC. Copy the new weights into the running pods and restart the Triton process in each one
• DD. Set the Deployment strategy to Recreate so all pods load the new version together

In LoRA, the low-rank update BA is multiplied by a constant before it is added to the frozen weight output. Which factor is used, where r is the rank and alpha is the lora_alpha hyperparameter?

• AA. square root of r
• BB. alpha / r
• CC. alpha x r
• DD. r / alpha

A fine-tuning job on 8 GPUs shows GPU utilization that alternates between 95% and 0% every few hundred milliseconds. An Nsight Systems trace shows the GPUs idle at the start of each step while CPU threads tokenize and collate the next batch. Which change is MOST likely to improve throughput?

• AA. Increase the number of dataloader workers and enable prefetching with pinned memory
• BB. Raise tensor parallelism from 2 to 4 to spread each layer over more GPUs
• CC. Enable activation checkpointing to lower memory use per step
• DD. Switch the optimizer states from FP32 to BF16 to shorten the update