NVIDIA GenAI LLM Associate Practice Test

The strongest response is the one about minimizing sensitive data exposure and applying privacy controls, because prompt text can absolutely contain regulated information. Performance and convenience do not override privacy obligations.

In prompt tuning, virtual (continuous) trainable prompt tokens are added only to the input embedding sequence. In prefix tuning (Li & Liang), virtual keys and values are prepended to the Query-Key-Value self-attention computations at *every* layer of the Transformer stack. Consequently, prefix tuning has a higher capacity for adaptation but introduces slightly more parameter overhead than prompt tuning.

Traditionally, KV cache for a sequence is allocated in contiguous GPU memory blocks. Because sequence lengths are unpredictable, engineers had to over-allocate memory for the maximum possible length (e.g., 2048 tokens). This resulted in massive memory waste (internal fragmentation) and prevented high batch sizes. PagedAttention (inspired by virtual memory paging in OS) allocates KV cache in small, fixed-size blocks (pages) in non-contiguous memory, mapping physical pages to virtual blocks as needed. This reduces fragmentation to near zero, enabling significantly larger batch sizes.

A production AI system needs measurable signals about quality, safety, speed, and user impact. Without those metrics, the team cannot tell whether the assistant is improving, regressing, or creating unacceptable risk.

The need to connect the model to current enterprise knowledge is correct because the one about connecting the model to current enterprise knowledge fits because pretraining alone does not guarantee fresh or company-specific answers. The cited source, NVIDIA NeMo Retriever, supports this answer for the LLM Fundamentals scenario rather than the adjacent distractors.

NeMo Retriever is correct because NeMo Retriever is built for document extraction, embedding, reranking, and related retrieval workflows that support grounded generative applications. The cited source, NVIDIA NeMo Retriever, supports this answer for the RAG and Knowledge Integration scenario rather than the adjacent distractors.

Use guardrails to evaluate inputs and outputs around the inference request is correct because to use guardrails to evaluate inputs and outputs around the inference request fits because policy risk can appear on either side of the interaction. The cited source, About Guardrails, supports this answer for the Safety, Governance, and Responsible AI scenario rather than the adjacent distractors.

Top-P (nucleus sampling) is correct because Nucleus sampling is controlled by the Top-P parameter. The cited source, NVIDIA NIM Documentation: Inference Parameters, supports this answer for the Prompting and Adaptation scenario rather than the adjacent distractors.

Speculative Decoding uses a small, fast model (the draft model) to generate several consecutive candidate tokens autoregressively. Since the draft model is small, this happens very quickly. Then, the primary, large model (the target model) processes all these candidate tokens in parallel in a single forward pass to determine which tokens are acceptable according to its own probability distribution. This maintains the exact mathematical output quality of the large model while achieving a 2x-3x speedup in latency.

Shadow testing (or shadow deployment) copies incoming real-world request traffic and routes it to the candidate model (the shadow model) in the background. The user still receives the response generated by the active production model. This allows the engineering team to collect realistic performance metrics (latency, throughput, output quality, error rates) under real load without any risk to active customers.