NVIDIA-Certified Associate: Generative AI LLMs (NCA-GENL) practice questions.

A team is generating customer-service responses with an LLM and wants more consistent output for factual questions. Which sampling parameter set is most appropriate?

A RAG system retrieves five chunks per query and stuffs them into the prompt. Users report that the model often ignores the retrieved context and falls back to its training-data knowledge. Which mitigation is most likely to help?

A team needs to fine-tune a 70B-parameter base model to follow a new product's instruction style. They have a single 8x H100 node available. Which fine-tuning approach is most feasible on that hardware?

An ML engineer needs to serve multiple LLMs from a single inference endpoint with automatic batching across concurrent requests. Which NVIDIA software is the canonical fit?

TensorRT-LLM applies several inference optimizations to LLM serving on NVIDIA GPUs. Which optimization keeps the GPU continuously busy by adding new requests to a running batch as old requests complete?

A production LLM serving cluster is constrained by KV-cache memory at high concurrency. Which technique most directly addresses the KV-cache bottleneck?

An ML engineer needs to add safety guardrails to a customer-facing LLM application: content moderation, off-topic refusal, jailbreak detection. NVIDIA ships an open-source toolkit specifically for this use case. Which one is the canonical fit?