Free NCA-GENM practice test: 30 questions with answers

Triton serves models across frameworks, with features such as dynamic batching and concurrent model execution for high throughput. DALI accelerates data preprocessing, Omniverse is for 3D simulation, and the CUDA compiler does not serve models.

TensorRT builds optimized inference engines through layer fusion, kernel selection, and reduced-precision execution. cuDF and RAPIDS handle dataframes and analytics, and Riva sample projects are not an inference optimizer.

Lower precision shrinks memory and speeds computation, and calibration helps limit accuracy loss. It does not increase parameter count, enlarge training data, or remove GPU needs.

The KV cache stores attention keys and values from previous tokens so they are not recomputed at each step, which speeds generation. Checkpoints, schedules, and augmentation are training-related concepts.

NIM packages optimized models as containers with standard APIs for GPU deployment. A printed specification, screenshot tool, and spreadsheet macro do not provide deployable inference services.

NeMo provides tools for building and customizing generative models across language, speech, and multimodal tasks. Jetson hardware guides, Omniverse Kit for simulation, and cuML for classical machine learning do not cover that scope.

Batching processes more requests per GPU pass, which raises throughput, but requests may wait briefly, adding latency. It does not reduce both metrics or guarantee no errors.

Efficient attention kernels and KV caching reduce the work per decoding step, which lowers time per token. Dataset size affects training, logs and fonts do not affect generation speed.

Vision-language models encode images and text into a shared or aligned representation so they can reason across both. Text-only models lack an image encoder, and regression or clustering models do not process images with language.

Vision transformers split images into patches and encode them as embeddings, which can be combined with text tokens. Spreadsheet tokenizers, loss functions, and schedulers do not produce image embeddings.

Contrastive training aligns images and captions in a shared space by rewarding matched pairs and penalizing mismatches. It does not generate images, compress audio, or count objects by itself.

Automatic speech recognition converts audio to text, which can then feed a language model. Segmentation, image generation, and anomaly detection do not process speech.

Motion depends on the order and timing of frames, so preserving temporal structure is essential. Single frames, discarded timing, and file names do not capture motion.

Multimodal models need an encoder for images and a tokenizer for text to convert both into model inputs. Printer drivers, spreadsheet engines, and DNS resolvers are unrelated to model architecture.

Aligned embedding spaces let different modalities be compared, which enables cross-modal search. Embeddings from different modalities are not identical, they are stored as vectors rather than images, and embedding size affects storage and compute.

A test set with brand colour checks measures compliance with the specific requirement. Prompt counts, microphone quality, and server racks do not test the output.

Diffusion models learn to reverse a gradual noising process, producing images by iterative denoising. Decision trees, PCA, and SVMs are not generative image models of this kind.

A VAE encodes inputs into a probabilistic latent space and decodes samples from it, enabling generation. Hand-written rules, fixed labels, and identity mappings do not learn a generative latent space.

In a GAN the generator tries to fool the discriminator, which learns to tell real from generated samples. Encoders that never interact, tokenizers and schedulers, and databases are not the GAN pair.

Self-attention provides direct connections between positions and supports parallel computation, which helps with long-range dependencies. Processing one token at a time describes older recurrent designs, and ignoring order or needing no data is incorrect.

Self-attention alone is order-agnostic, so positional encodings provide information about token positions. Compression, encryption, and language choice are different functions.

Decoder-only models are trained autoregressively to predict the next token from earlier tokens. Single-step full outputs, only class labels, and pixel averages are different tasks.

LoRA trains small low-rank adapter matrices while keeping most pretrained weights frozen, which reduces compute and memory. Training from random weights needs far more data and compute, and deleting weights removes the pretrained knowledge.

Mismatched pairs teach the model incorrect associations, so filtering or correcting them improves alignment. Training anyway, removing captions, and colour changes do not address the mismatch.

FID compares feature statistics of generated and real images, with lower values indicating closer distributions. Caption error, prompt length, and GPU counts are not distribution-level image metrics.

Prompt faithfulness is measured by alignment between text and images, using automatic scores and human review. File sizes, counts, and colour depth do not indicate whether the image matches the prompt.

Removing identifiers and following approved data-use processes protects patients and meets regulatory requirements. Public publication, partial removal, and personal storage expose sensitive data.

A held-out validation set shows whether the model generalizes beyond the training data, which reveals overfitting. Evaluating only on training data hides it, and removing validation or training without checks makes overfitting harder to detect.

Deduplication prevents over-representation and removing unsafe captions improves safety and quality. Keeping duplicates, dropping captions, and ignoring resolution each weaken the dataset.

Output safety classifiers and filters screen generated content before it reaches users, which addresses unsafe output. Batch size, tokenizers, and unreviewed storage do not screen outputs.