NVIDIA Certified Associate - Generative AI Multimodal practice questions

Question 1

A team is migrating a Crossfilter dashboard to cuxfilter to handle a much larger dataset on a GPU box, and a developer asks where the groupby, sort, and query operations will actually execute. Per the cuxfilter docs, how does cuxfilter process those operations?

Accepted Answer

It keeps the data as a GPU DataFrame and runs groupby, sort, and query on the GPU.

Answer

It copies the data into a Pandas DataFrame and runs the operations on CPU threads.

Answer

It streams every row to the browser and lets JavaScript compute the aggregations.

Answer

It offloads every operation to a remote Spark cluster and caches the charts locally.

Question 2

A data science team is explaining XGBoost's tree ensemble to stakeholders, and an engineer is asked how a CART leaf differs from an ordinary decision-tree leaf per the XGBoost docs. Which statement is correct?

Accepted Answer

A CART associates a real score with each leaf, giving richer meaning than a decision tree whose leaves hold only decision values.

Answer

A CART stores only a class label at each leaf, so it carries strictly less information than an ordinary decision-tree leaf.

Answer

A CART is restricted to classification targets, so its leaves cannot hold numeric scores and the ensemble outputs labels only.

Answer

A CART replaces each leaf with a separate linear regression model, which makes the boosted ensemble behave just like a random forest.

Question 3

A company is building a NeMo Curator video pipeline and must pick the component that detects transitions between shots so raw footage can be split into scene-aligned clips. Which model performs that shot-boundary detection, per NVIDIA's clipping documentation?

Accepted Answer

TransNetV2, a shot-boundary detection model that identifies transitions between shots.

Answer

TensorRT-LLM, which segments the stream by predicting per-frame caption boundaries first.

Answer

NV-CLIP, which scores frame-to-frame aesthetics and cuts on low-quality segments.

Answer

Cosmos Tokenizer, which compresses frames and emits clip boundaries as tokens.

Question 4

A platform team must choose between self-managing open-source inference engines and adopting NVIDIA NIM for LLMs, and it lacks the bandwidth to track the fast-moving LLM ecosystem. Per the NIM LLM overview, what does NIM LLM provide to address that gap?

Accepted Answer

Validated containers, curated weights, and direct alignment with upstream inference engines

Answer

A no-code GUI that fine-tunes any open model end to end without writing inference or training code

Answer

A billing proxy that resells third-party API tokens while leaving engine upgrades to the operator

Answer

A managed Kubernetes distribution that replaces the cluster's scheduler and ingress controllers

Question 5

A developer is running a chained .groupby().rolling().apply() expression through cudf.pandas, and the operation fails partway down the chain. How does cudf.pandas specifically handle this chained-call failure?

Accepted Answer

It rewinds and replays the chain minimally to deliver the correct result

Answer

It discards the whole chain and re-executes every step from scratch on the CPU

Answer

It forces every chained call in the session to run only on pandas afterward

Answer

It returns the last successful intermediate frame and drops the remaining steps

Question 6

A team is hitting a wall on per-GPU memory while deploying a model whose parameters exceed a single GPU's capacity, and is considering NeMo Model Parallelism. According to the NeMo parallelisms guide, what is Model Parallelism's primary purpose?

Accepted Answer

It partitions the model parameters across multiple GPUs to reduce the per-GPU memory needed to hold the model.

Answer

It replicates the full set of model parameters onto every GPU so each device can serve incoming requests independently.

Answer

It splits only the input data batch across the GPUs while keeping a complete copy of the model resident on each GPU.

Answer

It offloads idle model parameters to host CPU memory so that the GPUs only ever hold the currently active layer.

Question 7

An MLOps team is rolling out a NeMo Guardrails topic-control rail to keep a banking assistant on approved subjects, and they want to customize the wording of the restriction that forces the safety model to answer only "on-topic" or "off-topic". According to the documentation, what is the supported way to customize that output restriction?

Accepted Answer

Modify the TOPIC_SAFETY_OUTPUT_RESTRICTION variable in the topic_safety_check_input action.

Answer

Override the self_check_input prompt that drives the LLM self-check input rail's decision.

Answer

Edit the $relevant_chunks context variable that the self-check fact-checking rail consumes.

Answer

Register a new YARA injection rule within the agentic-security injection-detection input flow.

Question 8

An MLOps team is evaluating NeMo Curator's video pre-processing for a multimodal training experiment that historically took days on un-optimized GPU code. What concrete speed-up does NVIDIA's NeMo blog quote for NeMo Curator video preprocessing?

Accepted Answer

NeMo Curator streamlines visual data preprocessing, achieving up to 7x faster processing of video compared to unoptimized GPU implementations through multi-GPU load balancing across pipeline stages

Answer

NeMo Curator is CPU-only and is consistently slower than the same un-optimized GPU pipeline by about 10x; inside the documented NVIDIA reference architecture for this workload

Answer

NeMo Curator delivers approximately a 1.05x speedup over un-optimized GPU code, mostly from a faster JSON parser; for the data scale and latency budget the organization is targeting

Answer

NeMo Curator runs only on a single GPU and offers no multi-GPU load balancing benefit; across the GPU class, batching profile and storage tier currently provisioned

Question 9

A team is rolling out tensor parallelism across multiple GPUs to host a single very-large LLM. Which TensorRT-LLM mechanism aligns with that goal per NVIDIA's documentation?

Accepted Answer

Tensor parallelism where individual weight matrices are split across devices, enabling efficient inference at scale with each model running in parallel across multiple GPUs connected through NVLink and across multiple servers without developer intervention

Answer

Pipeline-only parallelism where each layer runs on a separate node and no weight matrix is ever split, forcing serialized execution; under the multimodal training pipeline the team has standardized on

Answer

Disabling NVLink to keep traffic on PCIe-only links, which is how TensorRT-LLM achieves multi-GPU performance; across the GPU class, batching profile and storage tier currently provisioned

Answer

A pure CPU sharding scheme that does not use GPUs and does not require any inter-device communication primitive; when the platform team applies the documented configuration in production

Question 10

A team is rolling out a custom video foundation-model training pipeline and needs to convert raw video into short, continuous clips before downstream curation. Which stage from the NVIDIA NeMo video-pipeline blog handles that step?

Accepted Answer

The clipping pipeline starts with decoding and splitting raw videos into short, continuous clips by analyzing frame-to-frame color changes

Answer

The pipeline only supports continuous streaming and does not produce discrete clip outputs; given the cluster, dataset and modality combination the team has

Answer

The team must hand-segment every video into clips because the pipeline does not include any automated clip-boundary detection

Answer

The clipping stage is purely an audio-only operation and ignores video frames entirely; across the GPU class, batching profile and storage tier currently provisioned

Question 11

A team is processing a dataset that is far larger than a single GPU's memory and needs out-of-core execution that can still scale across a multi-GPU cluster. According to Dask cuDF's documentation, how is this achieved?

Accepted Answer

Dask cuDF streams partitions to one GPU for out-of-core work, scaling via Dask-CUDA.

Answer

cuDF caches the full dataset in host RAM and pages individual columns onto one GPU.

Answer

cuML's fit-predict API shards the dataset and trains one model replica per GPU.

Answer

Triton ensembles split the dataset and run each split as a concurrent model instance.

Question 12

Before production, an analyst is choosing how to handle class imbalance in a tabular ML pipeline and wants the NVIDIA-blessed list of mitigation techniques for feature engineering. Which combination matches NVIDIA's Spark-RAPIDS data-preprocessing overview?

Accepted Answer

Class imbalance mitigation (oversampling with SMOTE, undersampling, class weights)

Answer

Class imbalance is mitigated only by shrinking the test set so that minority and majority classes appear equal at evaluation time

Answer

Class imbalance is never addressed in tabular ML and is always considered acceptable by NVIDIA's guidance

Answer

Class imbalance is corrected exclusively by relabeling minority-class rows to the majority class

Question 13

A company is migrating an LLM to TensorRT-LLM and needs to cut memory footprint and compute cost by converting BF16 weights to a lower-precision type. According to the TensorRT-LLM quantization docs, which statement correctly characterizes the available quantization recipes?

Accepted Answer

Quantization converts weights and/or activations to lower-precision types like INT8, FP8, or FP4, via recipes such as W4A16 AWQ.

Answer

Quantization in TensorRT-LLM converts only activations and never weights, and the one supported target type is plain INT8.

Answer

Quantization first upconverts the BF16 weights to FP32, then exposes only the single NVFP4 KV Cache recipe for inference workloads.

Answer

Quantization is restricted to the FP8 KV Cache recipe and cannot be applied to model weights or to dense activation tensors.

Question 14

A team has decided to layer NeMo Guardrails across an LLM application that needs BOTH input sanitization and post-hoc output filtering for unsafe content. Which TWO NeMo Guardrails rail types from the documented five-stage taxonomy correspond to those two responsibilities?

Accepted Answer

Input Rails validate and sanitize user inputs before the LLM is called to prevent harmful content from reaching the model

Accepted Answer

Output Rails provide final protection by evaluating model responses, filtering, editing, or blocking unsafe or off-policy content before users receive them

Answer

A separate code-review rail that lints application source code statically before the LLM is shipped; when the platform team applies the documented configuration in production

Answer

A purely UI-only banner that warns users to be careful before any LLM call is made; when the platform team applies the documented configuration in production

Question 15

A team is evaluating a RAG pipeline in NVIDIA NeMo Microservices and must score both retrieval recall AND whether generated answers stay factually consistent with the retrieved context. Which TWO metric families documented for NeMo Microservices RAG evaluation cover those two requirements?

Accepted Answer

Recall@k (fraction of relevant documents in top k results) for retrieval quality

Accepted Answer

Faithfulness (factual consistency of answer vs context, 0-1 scale, requires judge LLM) for grounded answer evaluation

Answer

Average request-latency in milliseconds measured by the NIM autoscaler; for the data scale and latency budget the organization is targeting

Answer

Embedding-store disk usage in gigabytes as a proxy for answer correctness; inside the documented NVIDIA reference architecture for this workload

NVIDIA Certified Associate - Generative AI Multimodal

Sample questions

Sources

Similar exams