Hardest NVIDIA-Certified Professional Generative AI LLMs questions

Multi-head attention works because each head has independent linear projections (Q, K, V) for query, key, and value vectors. These projections are randomly initialized and learn different representations during training. This allows heads to specialize: some learn to attend to adjacent tokens (useful for syntax), others to distant tokens (useful for semantics). No explicit constraint or manual role assignment is needed—specialization emerges from the learning dynamics. The transformer does not use separate layers for local vs. global patterns; all heads operate in parallel. Heads are not manually assigned roles or given different cache sizes. While each head operates on a subset of the embedding dimension, the optimizer does not force specific local/global role assignments—this is an emergent property.

The Guardrails jailbreak protection page warns that jailbreak detection fails open. When the configured detector cannot be reached, the rail logs the failure, allows the request, and returns no error to the caller, so an outage silently removes the protection. The page advises monitoring detector availability and pairing this rail with rails that do not call an external service. Option A is the reverse of the documented behavior. Option B describes queuing that is not documented. Option D is wrong because in-process checks run only when server_endpoint is unset, and the page says that mode is for testing only.

Chain-of-thought reasoning combined with constrained JSON decoding is the best approach. Chain-of-thought makes the recommendation logic transparent and improves quality. Constrained decoding enforces the exact schema (product ID, reason, confidence score) at generation time, guaranteeing valid, structured output. Few-shot with greedy decoding improves consistency but does not enforce structure. Zero-shot with post-processing is fragile. One-shot alone is insufficient for reliable recommendations.