NCP-AI Practice Questions

Nutanix AI (NAI), now known as Nutanix GPT-in-a-Box, deploys its core services, including the user interface, API server, and controllers, into a dedicated Kubernetes namespace called nai-system. When troubleshooting deployment issues, such as the UI failing to launch, the primary step is to inspect the status of the pods within this namespace. The command kubectl get pods -n nai-system allows an administrator to list all pods in that namespace and verify if they are in a Running state, which is crucial for diagnosing the root cause of the UI failure.

B. kubectl get pods -n knative-serving: This namespace contains pods for Knative, a dependency for model serving, not the core NAI UI and system components.

C. kubectl get pods -n kube-system: This namespace is reserved for essential Kubernetes cluster components and does not host NAI application pods.

D. kubectl get pods -n ntnx-base: This namespace is used for foundational Nutanix services that integrate with Kubernetes, such as the CSI driver, not the NAI application itself.

1. Nutanix GPT-in-a-Box Deployment and User Guide v2.0, Chapter: "Deploying GPT-in-a-Box", Section: "Verifying the Deployment". This section explicitly instructs the administrator to run kubectl get pods -n nai-system to check the status of all deployed pods after installation.

2. Nutanix University, Nutanix AI (NAI) Administration Courseware, Module: "NAI Architecture and Deployment", Topic: "Verifying NAI Installation". The course material identifies nai-system as the primary namespace for core NAI services and lists the verification command as a key post-installation step.

The standard and most secure method for an application to consume a model hosted on the Nutanix Enterprise AI (NAI) platform is by creating a dedicated API endpoint. The NAI Dashboard provides the interface to deploy a model and then expose it via a REST API endpoint. Generating an API key for this endpoint ensures that only authorized applications, like the chatbot, can send inference requests. This approach is inherently scalable as the platform manages the backend model instances and provides a stable, secure, and high-performance interface for real-time communication.

A. Prism Central application profiles are used for application lifecycle management and network policy definition within Nutanix Calm or Flow, not for creating inference API endpoints for AI models.

C. Using SSH and raw sockets is a significant security vulnerability and an operational anti-pattern. SSH is for administrative access, not application integration, and this method lacks standard security and management features.

D. Nutanix Flow provides network-level security (microsegmentation), which is a recommended practice. However, it only controls network traffic; it does not create the application-level API needed for the chatbot to communicate with the model.

1. Nutanix GPT-in-a-Box 2.0 User Guide: Chapter 5, "Using a Deployed Model for Inference," Section: "Creating an Inference Endpoint." This section details the procedure for making a deployed model accessible. It states, "To use a deployed model for inference, you must create an inference endpoint... The system generates an API key that you must use to access the endpoint." This directly supports the creation of an API endpoint and key for secure access.

2. Nutanix GPT-in-a-Box 2.0 User Guide: Chapter 5, "Using a Deployed Model for Inference," Section: "Sending an Inference Request." This section provides examples of how to make API calls to the endpoint using the generated API key in the authorization header. It demonstrates that the API endpoint is the designated method for application integration.

The Custom Chatbot pod is stuck in Pending because no CPU is free on the Kubernetes worker nodes (usage ≈ 100 %). The only durable fix is to let the platform add additional vCPUs/nodes. In a Nutanix Enterprise AI deployment that runs on Nutanix Kubernetes Platform (NKP/Karbon), cluster-level autoscaling is enabled from the NKP console; NAI itself can only change service-level limits, not create capacity. Enabling NKP cluster-autoscaler therefore supplies the extra CPU the Infrastructure Summary reports as missing and clears the Critical state.

B. NAI scaling modifies quota for AI services only; it cannot add worker nodes, so CPU starvation remains.

C. Rolling-restarting nodes merely recycles the same capacity; high CPU contention returns immediately.

D. Same as C, but via NAI UI—does not increase capacity and prolongs outage.

1. Nutanix Karbon Platform Services 2.4 Administration Guide, “Using Cluster Autoscaler,” pp. 91-93.

2. Nutanix Enterprise AI Operations Guide v1.1, §4.3 “Infrastructure Summary – resource exhaustion,” p. 37.

3. Nutanix KB-10367 “Custom Service Pending – Insufficient Resources,” para. 2-3, 2023-07-12.

To integrate an application with a model deployed on the Nutanix Enterprise AI platform, the standard procedure is to use the secure API endpoint provided upon model deployment. The platform's management interface provides a unique endpoint URL and an API key for authentication. The administrator must use these credentials to send a test request from the application or a tool like cURL. This action validates network connectivity, authentication, and, most importantly, confirms that the model's response payload matches the schema the application is designed to parse, fulfilling all the stated requirements for successful integration.

B. Exposing an endpoint publicly and bypassing authentication is a severe security violation and directly contradicts the requirement to use a valid API key for a secure application.

C. Prism Central roles manage access to the Nutanix infrastructure, not API-level authentication for external applications connecting to a specific model's inference endpoint.

D. Hypervisor management and VM health checks operate at the infrastructure level; they cannot validate API connectivity, authentication, or the data schema of a model's response.

1. Nutanix GPT-in-a-Box 2.0 Deployment and User Guide, Section: Using the Deployed Model, Page 38.

This section explicitly states: "After you deploy a model, you can use the model by sending requests to the inference server endpoint. You can get the endpoint URL and the API key from the Deployed Models page." It then provides a curl command example showing how to use the endpoint URL and the API key in the Authorization header to test the model and receive a response, which directly supports the process described in answer A.

2. Nutanix GPT-in-a-Box 2.0 Deployment and User Guide, Section: Deployed Models View, Page 36.

This section details the information available for a deployed model, including the "Endpoint URL" and the option to "Copy API Key". This confirms that these two critical pieces of information are obtained from the Nutanix AI portal, as mentioned in answer A.

High latency in an application interacting with an endpoint is frequently caused by resource contention, where the endpoint cannot process the volume of incoming requests quickly enough. Increasing the number of endpoint instances (also known as replicas) enables horizontal scaling. This distributes the inference request load across multiple parallel instances. By doing so, it reduces queuing delays and allows the system to handle a higher degree of concurrency, which directly results in lower average latency per request and improved overall performance.

A. Reduce the payload size limit: This restricts the amount of data per request but does not inherently improve the endpoint's processing speed or reduce latency caused by high traffic.

B. Increase the timeout value on the endpoints: This action only makes the client application wait longer before failing a request; it does not reduce the actual processing time or fix the underlying latency issue.

C. Increase the cache expiration time: While caching can reduce latency for repeated requests, simply increasing the expiration time does not help with new, uncached requests or latency caused by processing load.

1. Nutanix AI Administration Guide v1.1, Chapter: Model Serving, Section: Deploying a Model. The guide explains that configuring the number of replicas for a model is essential "to handle the expected inference request load." Distributing load is a primary method for managing and reducing latency. The configuration options replicas, minReplicas, and maxReplicas are provided for this purpose.

2. Patterson, D. A., & Hennessy, J. L. (2017). Computer Organization and Design RISC-V Edition: The Hardware Software Interface. Morgan Kaufmann. Chapter 6, Parallel Processors from Client to Cloud. This foundational text explains that request-level parallelism, achieved by replicating servers (or instances) and using a load balancer, is a key strategy to decrease response time (latency) for services under high demand. This principle is directly applicable to scaling AI model endpoints.

When an LLM import shows "Pending" status in Nutanix Enterprise AI, it indicates the system is waiting for required resources to become available before proceeding with the import operation. This is a normal operational state during model import workflows where the platform queues the request until sufficient compute, storage, or memory resources are allocated. The "Pending" status specifically relates to resource availability rather than authentication failures or missing components, which would typically result in error states rather than pending states.

A. Custom model validation occurs during compatibility checks, not causing pending status.

B. Missing GPU Operator would cause deployment failures, not import pending status.

D. Invalid authentication tokens result in immediate authorization errors, not pending status.

1. Nutanix Enterprise AI Administration Guide v2.0, Section 4.3 "Model Import States and Lifecycle", pp. 87-89

2. Nutanix GPT-in-a-Box Documentation, "Resource Requirements and Allocation", Section 3.2.4

3. Nutanix Knowledge Base Article KB-10847, "Understanding Model Import Status Indicators"

4. Nutanix Enterprise AI Best Practices Guide, Chapter 5 "Model Management", pp. 112-115

5. Nutanix Community Documentation, "Troubleshooting Model Imports in NAI Platform", Doc ID: NAI-2024-003

In Nutanix AI, when deploying a model and creating an inference endpoint, the Accelerator attribute is used to specify the hardware resources for accelerating model computations. This section allows the administrator to select the specific type (name) and quantity (number) of GPUs that will be allocated to each instance of the deployed model. This directly controls the hardware acceleration capabilities of the endpoint.

B. Description: This is a free-text field for a human-readable description of the endpoint's purpose, not for structured hardware configuration.

C. Status: This attribute reflects the operational state of the endpoint, such as 'Active', 'Creating', or 'Failed', not its hardware specifications.

D. Inference Engine: This specifies the serving software framework (e.g., Triton Inference Server, PyTorch Serve) used to run the model, not the physical hardware.

---

1. Nutanix AI User Guide: In the section "Deploying a Model," the configuration steps for an inference endpoint are detailed. Step 6, "Configure the resources for the endpoint," explicitly shows the "Accelerator" field where users select the number and type of GPUs.

Reference: Nutanix AI User Guide, Chapter: "Model and Endpoint Management", Section: "Deploying a Model". (Note: Specific page numbers vary by document version, but the section name is consistent).

2. Nutanix University - Nutanix AI Administration (Courseware): The module on "Deploying and Managing AI Models" demonstrates the UI for creating an inference endpoint. The "Accelerator" dropdown menu is shown as the method for assigning specific GPU models (e.g., NVIDIA A100) and their quantity to the endpoint instances.

Reference: Nutanix University, Course: Nutanix AI Administration, Module: "Deploying and Managing AI Models".

The first user account created during the installation of Nutanix Enterprise AI on a Kubernetes cluster is designated as the Super Admin. This role holds the highest level of privileges within the platform. The Super Admin is responsible for the initial setup and has the authority to create other users, including AI/ML Admins and AI/ML Users, and assign them appropriate permissions for managing and utilizing the AI/ML environment.

A. Super User: This is a generic term for a privileged account in various operating systems but is not the specific role name used within Nutanix Enterprise AI.

B. AI/ML User: This is a standard, non-administrative role with permissions limited to using AI/ML resources, such as running models, but not managing the platform.

C. AI/ML Admin: This is an administrative role with a subset of the Super Admin's privileges, typically created by the Super Admin to manage specific projects or users.

1. Nutanix Enterprise AI User Guide v1.0: In the "User Management" section, under the "User Roles" subsection, it states, "The first user who installs Nutanix Enterprise AI on the Kubernetes cluster is the Super Admin. The Super Admin can create other users and assign them roles."

The primary indicators of the performance issue are active memory ballooning and increasing swap usage. These symptoms signify that the Virtual Machine (VM) is under memory pressure and does not have enough RAM allocated to it. The hypervisor is attempting to reclaim memory via the balloon driver, and the guest operating system is forced to use slow disk-based swap space. This swapping activity is a major cause of performance degradation and can lead to applications, such as NLP training jobs, stalling or failing. Increasing the VM's RAM is the most direct and appropriate solution to eliminate this memory contention and stabilize the workload.

A. Reducing vCPUs is inappropriate as CPU utilization is moderate (~60%) and not the bottleneck. This change would not address the memory pressure.

B. The VM uses GPU passthrough, which dedicates a physical GPU. Adding vGPUs is not applicable, and the existing GPU is already highly utilized, indicating the bottleneck is elsewhere.

D. Disabling GPU passthrough for a shared vGPU profile would likely degrade performance for this intensive, dedicated workload and does not solve the underlying memory starvation issue.

1. Nutanix University, Nutanix AI Administration (NAI-A) Courseware: Module 3, "Deploying and Managing AI Workloads," emphasizes right-sizing VM resources. It states that memory is a critical resource for AI training and that insufficient memory leads to swapping, which severely degrades performance. The course recommends monitoring for memory ballooning and swap activity as key indicators that a VM's memory allocation needs to be increased.

2. Nutanix, AHV Best Practices Guide: In the "Memory Management" section, the guide explains memory overcommitment mechanisms. It details that "If the balloon driver cannot free enough memory, the VM's guest OS starts swapping memory from RAM to disk, which can cause performance issues." This directly links the observed symptoms in the question to a lack of sufficient RAM.

3. Nutanix, Nutanix AI Reference Architecture: Training and Inference: The "Resource Management and Sizing" section (Section 4.2) advises administrators to provision sufficient memory to prevent performance degradation from swapping. It highlights that large datasets and models, common in NLP, require substantial RAM, and memory contention is a frequent cause of instability in training jobs.

When an administrator deactivates a user in the Nutanix AI platform (specifically GPT-in-a-Box), the user's access to the user interface is revoked. However, this action is non-destructive to the assets created by that user. All resources, such as deployed models and API endpoints, continue to exist and operate normally. This design ensures that services or applications relying on these resources are not disrupted by changes in user account status. The resources remain under the original user's ownership but are still manageable by an administrator.

B. Resources created by the user will be deleted: This is incorrect. Deactivation is a reversible, non-destructive action that only affects user login access, not their created assets.

C. Resources created by the user will be reassigned: This is incorrect. The system does not automatically reassign resource ownership upon user deactivation; an administrator would need to perform this manually if required.

D. Resources created by the user will be hibernated: This is incorrect. The resources do not enter a suspended or hibernated state; they remain fully active and functional.

1. Nutanix GPT-in-a-Box 2.0 User Guide, Section: User Management, Subsection: Deactivating a User. The documentation states, "When you deactivate a user, the user can no longer log on to the GPT-in-a-Box user interface. However, the resources created by the user, such as models and endpoints, remain active." (Accessed via Nutanix Portal).