Free NVIDIA Generative AI Multimodal Certification Exam Practice Questions

Sample Practice Questions

Try these 10 sample questions from the NVIDIA Generative AI Multimodal Certification Exam practice bank. Each includes the answer and explanation. Continue with the interactive practice above to review the full question set.

1. 1A multimodal training loop uses PyTorch GradScaler and clips gradients with a threshold intended for unscaled gradients. After backward and before clipping, what should the team do?

   1. A.Unscale the optimizer's gradients.Correct answer: A
   2. B.Update the scaler's scale factor.
   3. C.Run the optimizer step immediately.
   4. D.Clear the optimizer's gradients.

   Explanation: PyTorch states: "Calling scaler.unscale_(optimizer) before clipping enables you to clip unscaled gradients as usual:" Updating, stepping, or clearing does not prepare the accumulated gradients for this clipping operation.

   Reference: Automatic Mixed Precision examples — Working with Unscaled Gradients, Gradient clipping

2. 2An analyst uses scikit-learn KMeans to group an unlabeled dataset. Which parameter sets the number of clusters to form?

   1. A.n_clustersCorrect answer: A
   2. B.n_init
   3. C.max_iter
   4. D.tol

   Explanation: "The number of clusters to form as well as the number of centroids to generate." This is the meaning of n_clusters; n_init controls runs, max_iter limits iterations, and tol governs convergence.

   Reference: KMeans — scikit-learn — Parameters, n_clusters

3. 3In the DDPM training objective described in the Hugging Face course, what does the model predict?

   1. A.The original clean image itself.
   2. B.A velocity derived from image and noise.
   3. C.The noise used to corrupt the input.Correct answer: C
   4. D.The difference between two clean images.

   Explanation: "In DDPM and many other diffusion model implementations, the model predicts the noise used in the corruption process" The course distinguishes this noise target from predicting the clean image or a velocity.

   Reference: Diffusion Models from Scratch — Training Objective

4. 4Under a senior engineer's supervision, a developer wants Triton to batch independent inference requests for a stateless model. Which Triton batcher is intended for this use?

   1. A.The sequence batcher for stateful request sequences.
   2. B.The ensemble scheduler for model pipelines.
   3. C.Direct execution without server-side batching.
   4. D.The dynamic batcher for stateless requests.Correct answer: D

   Explanation: "The dynamic batcher should be used for stateless models." The sequence batcher targets stateful sequences that must remain on the same instance.

   Reference: Batchers — NVIDIA Triton Inference Server — Dynamic Batcher

5. 5For the same number of values, how does FP16 storage compare with FP32 storage in NVIDIA's mixed-precision guide?

   1. A.FP16 uses half as many bytes.Correct answer: A
   2. B.FP16 uses one quarter as many bytes.
   3. C.FP16 uses the same number of bytes.
   4. D.FP16 uses twice as many bytes.

   Explanation: "Speed up memory-limited operations by accessing half the bytes compared to single-precision." This comparison refers to FP16 versus single-precision FP32 values.

   Reference: Train With Mixed Precision — Automatic Mixed Precision benefits

6. 6During client scoping of an ML project, which discussion establishes both the intended use case and a way to judge its success?

   1. A.Define the model's task and agree on evaluation criteria.Correct answer: A
   2. B.Select the optimizer and defer defining success criteria.
   3. C.Estimate compute cost without defining the prediction target.
   4. D.Choose deployment hardware before defining the use case.

   Explanation: "Before building anything, align on what the model needs to do and how you will know it is working." The scope and success criteria guide later design and evaluation.

   Reference: Machine learning lifecycle — Scope the use case and define success

7. 7According to NIST's AI RMF, what does transparency describe?

   1. A.The capacity to resist unauthorized changes.
   2. B.The meaning of an output in its use context.
   3. C.The protection of individuals' information.
   4. D.The availability of information about an AI system and its outputs.Correct answer: D

   Explanation: "Transparency reflects the extent to which information about an AI system and its outputs is available" NIST also notes that transparency alone does not establish accuracy, privacy, security, or fairness.

   Reference: AI Risks and Trustworthiness — 3.4 Accountable and Transparent

8. 8A multimodal training loop uses PyTorch gradient norm clipping. Which setting causes the clipping call to raise an error when the total gradient norm is NaN or infinite?

   1. A.Set foreach to True.
   2. B.Set error_if_nonfinite to True.Correct answer: B
   3. C.Set error_if_nonfinite to False.
   4. D.Set norm_type to infinity.

   Explanation: The documentation states: "if True, an error is thrown if the total norm of the gradients from parameters is nan, inf, or -inf." The error_if_nonfinite option controls this check; norm_type selects the norm, and foreach selects an implementation.

   Reference: torch.nn.utils.clip_grad_norm_ — PyTorch documentation — torch.nn.utils.clip_grad_norm_, error_if_nonfinite parameter

9. 9Which criterion does scikit-learn KMeans minimize when choosing centroids?

   1. A.Maximum distance from a point to its centroid.
   2. B.Within-cluster sum of squared distances.Correct answer: B
   3. C.Between-cluster sum of squared distances.
   4. D.Sum of distances to the global dataset mean.

   Explanation: "The K-means algorithm aims to choose centroids that minimise the inertia, or within-cluster sum-of-squares criterion:" Inertia measures the distances of samples to their cluster centroids.

   Reference: Clustering — scikit-learn — 2.3.2 K-means

10. 10In the course's DDPM training example, which loss is computed between the model prediction and its noise target?

    1. A.Categorical cross entropy.
    2. B.Binary cross entropy.
    3. C.Hinge classification loss.
    4. D.Mean squared error.Correct answer: D

    Explanation: "loss = mse_loss(model_prediction, noise) # noise as the target" The example compares the predicted noise with sampled noise using MSE.

    Reference: Diffusion Models from Scratch — Training Objective