Free NCP-GENL practice test: 30 questions with answers

The KV cache keeps attention keys and values for previously processed tokens, so generation only computes the new token's contribution. Dropout, gradient accumulation and label smoothing are training-time techniques.

Continuous batching schedules requests dynamically so the GPU stays busy and throughput rises under concurrency. Serial processing and disabling the KV cache reduce throughput, and CPU-only execution is much slower for LLM inference.

FP8 uses one byte per value, halving memory compared with FP16 and supported natively on recent NVIDIA GPUs such as Hopper and Blackwell. FP64 doubles memory use, and unscaled integers or arbitrary decimals are not standard inference formats.

A causal mask sets attention scores for future positions to negative infinity, so each token attends only to itself and earlier tokens. Bidirectional cross-attention belongs to encoder-decoder models, and pooling or batch normalization do not restrict attention order.

RLHF and related preference-optimization methods adjust the model using ranked human judgments so outputs follow instructions and preferences. Warmup, checkpointing and tokenizer changes address optimization or input representation rather than alignment.

Dot products grow in magnitude with the key dimension, which pushes softmax into regions with tiny gradients. Dividing by the square root of the key dimension keeps variance stable. The other options do not address this scaling effect.

RAG retrieves current passages at query time and supplies them as context, so updating the index updates answers without retraining. A yearly snapshot goes stale, temperature changes randomness rather than knowledge, and a fixed prompt cannot hold a changing corpus.

An embedding model maps text to dense vectors that capture semantic similarity, which the vector index uses for retrieval. A reranker scores candidates after retrieval, a tokenizer splits text into tokens, and schedulers affect training only.

Reranking and careful context ordering with explicit grounding instructions make the relevant passage more likely to be used. Adding random documents dilutes context, removing metadata loses useful filters, and a smaller window reduces what can be supplied.

Reviewing training data, removing duplicated sensitive content, and filtering outputs address memorization and leakage directly, and a privacy evaluation confirms the result. The other changes affect model size or optimization, not leakage.

Guardrails and safety classifiers screen inputs and outputs against defined policies, which can block or modify harmful content. Gradient clipping, learning-rate decay and beam search are training or decoding techniques.

Safety tuning teaches appropriate refusals, and an output classifier catches harmful responses that slip through. Longer outputs, no system prompt and high-temperature sampling each increase risk.

Recall at k and ranking metrics such as mean reciprocal rank measure whether relevant passages were retrieved and where they appear. Generator perplexity, answer length and hardware count do not measure retrieval quality directly.

Groundedness or faithfulness metrics compare claims in the answer with the supplied evidence to detect unsupported statements. Formatting, word length and latency say nothing about factual support.

Instruction-following evaluations check responses against explicit, verifiable constraints such as length or format, which gives objective scores. Parameter counts, letter counts and temperature do not measure instruction following.

Tensor parallelism shards matrix multiplications across devices so each holds part of a layer's weights. Data parallelism replicates the full model, which does not solve the memory limit, and sweeps and early stopping do not split weights.

Pipeline parallelism places contiguous layer stages on different devices and pipelines micro-batches through them to keep devices busy. Full replication describes data parallelism, and tokenizer sharding and per-layer labels are not part of the technique.

ZeRO and fully sharded data parallel partition training state across workers so each GPU keeps only a slice. Feature hashing, dropout schedules and search methods do not partition memory.

Deduplication removes repeated or near-duplicate text so the model does not over-weight and memorize it. Batch size, context length and punctuation removal do not reduce repetition in the training data.

Byte-level BPE starts from raw bytes, so any Unicode text can be encoded, and learned merges compress frequent sequences. Word lookup produces unknown tokens for unseen words, while ASCII-only or hashed tables lose information.

LoRA freezes the pretrained weights and learns small low-rank adapters, which cuts trainable parameters and optimizer memory. It is used with decoder models as well, and it still relies on a normal optimizer.

QLoRA keeps the base model quantized to 4 bits and trains only LoRA adapters, which fits large models onto limited memory. Full float32 fine-tuning needs far more memory, random initialization discards pretraining, and training nothing does not adapt the model.

Chain-of-thought prompting elicits intermediate steps, which often improves accuracy on multi-step problems. Empty input, removing examples and one-word answers reduce the scaffolding that helps reasoning.

Keeping trusted instructions separate from untrusted input, validating what arrives, and checking outputs reduces prompt injection risk. Mixing roles, longer outputs and caching do not defend against injection.

TensorRT-LLM builds optimized engines for large language models, including fused attention and in-flight batching. cuDF and cuML target dataframes and classical machine learning, and Riva targets speech services.

Triton Inference Server hosts multiple models and frameworks with batching, versioning and GPU scheduling. The other products address cybersecurity analytics, optimization or data curation rather than model serving.

NeMo is NVIDIA's framework for training, customizing and aligning generative models, including tensor and pipeline parallelism. Omniverse targets simulation, Jetson is an edge hardware platform and Clara focuses on healthcare imaging.

NeMo Curator provides scalable pipelines for filtering, deduplication and curation of training data. Guardrails enforce conversation policies, Retriever supports retrieval pipelines, and the agent toolkit builds agent workflows.

Supervised fine-tuning on high-quality examples teaches consistent tone and structured output. A longer context does not change behavior by itself, and disabling the tokenizer or random selection break the pipeline.

Continued pretraining on unlabeled domain text adapts the model's knowledge and vocabulary usage, and instruction tuning can follow. Prompting alone does not change weights, and a tokenizer by itself does not teach facts.