NCP-AI Sample Questions Answers

Comprehensive and Detailed Explanation From Exact Extract of Nutanix Artificial Intelligence documents knowledge without any external URL or links:

Nutanix Enterprise AI utilizes Role-Based Access Control (RBAC) to manage permissions.6

AI/ML Admin:This role has full privileges over the NAI platform, including the ability tocreate new users, assign roles (Admin vs.7User), manage global quotas, and view all endpoints and models across the system.

AI/ML User:This role is restricted to managing only theirownresources (models, endpoints, and keys) and cannot create other user accounts or view resources belonging to other users.

Comprehensive and Detailed Explanation From Exact Extract of Nutanix Artificial Intelligence documents knowledge without any external URL or links:

Nutanix Enterprise AI provides an inference gateway that is API-compatible with the OpenAI standard, allowing developers to use standard libraries (like the OpenAI Python SDK) to interact with models running on NAI. The supported endpoints that map to this standard include:

/v1/chat/completions: For chat-based text generation (supported).

/v1/embeddings: For generating vector embeddings (supported).

/v1/models: For listing the available models (supported).

The option/v1/systemsisnota supported or valid OpenAI-compatible endpoint. In the OpenAI API schema, "system" refers to a messagerole(e.g., {"role": "system", "content": "..."}), not a standalone URL endpoint. Therefore, this is the unsupported type in the list.

Comprehensive and Detailed Explanation From Exact Extract of Nutanix Artificial Intelligence documents knowledge without any external URL or links:

According to the endpoint creation workflow in the Nutanix Enterprise AI documentation, an endpoint is essentially a deployment of a specific model. Therefore, the prerequisite for creating an endpoint is having a model available in the local registry.

Model Status:The system explicitly requires that the imported Large Language Model (LLM) be in theActive(or Ready) state before it can be selected in the "Create Endpoint" wizard.

Workflow:When an administrator imports a model (whether from Hugging Face or manually), it transitions through statuses like "Processing" (downloading) or "Pending" (waiting for resources). It is only when the status reachesActivethat the model is fully validated, stored, and available to be served as an endpoint. An "Uploaded" status typically refers to a manual file transfer that has not yet been processed or validated for deployment.

Comprehensive and Detailed Explanation From Exact Extract of Nutanix Artificial Intelligence documents knowledge without any external URL or links:

When verifying the installation of Nutanix Enterprise AI (NAI) via the kubectl get pods -n nai-system command, the READY column indicates the number of containers in the pod that are currently running and healthy compared to the total number of containers defined for that pod.

Meaning of 2/2:Critical system pods in the NAI architecture (such as monitoring components like prometheus-nai or pods equipped with service mesh/proxy sidecars) are often configured with multiple containers—typically the main application container and a sidecar (e.g., for metrics collection, logging, or traffic management).

Health Indication:A value of2/2confirms thatboththe primary container and its essential sidecar are fully up and running. This state indicates the pod is completely healthy and operating with its full high-availability or monitoring capabilities enabled. A value of 1/2 would indicate a degraded state where one container has failed, and 1/1 applies to single-container pods, but 2/2 is the specific indicator for multi-container system pods being fully operational.

Comprehensive and Detailed Explanation From Exact Extract of Nutanix Artificial Intelligence documents knowledge without any external URL or links:

To determine which versions of the nai-core Helm chart are available for installation from the repository, the administrator must use the helm search command. Specifically, the documentation states that after adding the Nutanix Helm repository, the command to search for the version of the nai-core helm chart available is helm search repo ntnx-charts/nai-core --versions.

helm get is used to retrieve information about a deployed release, not to search a repository.

helm pull is used to download the chart, not to list versions.

helm push is used to upload charts.

Comprehensive and Detailed Explanation From Exact Extract of Nutanix Artificial Intelligence documents knowledge without any external URL or links:

In the Nutanix Enterprise AI (NAI) endpoint lifecycle, the Pending status has a specific definition related to infrastructure resource availability. According to the endpoint management documentation:

"Pending: The system is waiting for the resources required to create the endpoint. For example, if the required number of GPUs are not available, Nutanix Enterprise AI maintains a Pending status until the GPUs become available."

The exhibit shows the administrator explicitly requested1 NVIDIA-H100-NVLGPU for the instance. If the underlying Kubernetes cluster does not have a free H100 GPU (e.g., all H100s are currently occupied by other active endpoints or training jobs), the scheduler cannot place the pod, and the status will remain "Pending" indefinitely until a GPU is freed.

In contrast:

Unauthorized NGC Keystypically result in aFailedstatus due to image pull errors (ImagePullBackOff) or authentication failures, not a resource-wait "Pending" state.

Permissions issueswould generally prevent the creation wizard from completing or result in an immediate 403 Forbidden error upon submission.

Wrong GPU selection(compatibility) is unlikely here as the "gemma-2-2b-it" is a small model that can easily run on the powerful H100; the issue is simply that the requested hardware is currently fully utilized.

Comprehensive and Detailed Explanation From Exact Extract of Nutanix Artificial Intelligence documents knowledge without any external URL or links:

The Nutanix Enterprise AI documentation provides distinct definitions for each import status state, which is crucial for troubleshooting.

Pending vs. Processing:

Processing:This status indicates that the system is activelydownloadingthe LLM or NIM from the external repository (Hugging Face or NGC).

Pending:This status specifically indicates that "The system is waiting for the resources required to save the LLM." The documentation explicitly cites"if the required storage space is not available"as the primary example for why a model would remain in thePendingstate.

Failed States:Options A (Read access) and B (Unauthorized token) would result in aFailedstatus, not Pending.

Conclusion:Since the status is "Pending" rather than "Processing," the issue is resource-related (storage), not the active download process.

Comprehensive and Detailed Explanation From Exact Extract of Nutanix Artificial Intelligence documents knowledge without any external URL or links:

TheEndpoint Detailspage in the Nutanix Enterprise AI dashboard contains specific widgets for monitoring the performance of a single inference service.4

Number of Instances Widget:This specific widget displays the scaling configuration and resource consumption of the endpoint. It explicitly lists theNumber of vCPUsand theMemory usage per instancethat is actively in use to serve the endpoint.5

Distinction:The "Infrastructure Summary" widget is located on the main Dashboard (not the Endpoint Details page) and displayscluster-wideresource usage, not the specific memory footprint of a single endpoint.

Comprehensive and Detailed Explanation From Exact Extract of Nutanix Artificial Intelligence documents knowledge without any external URL or links:

In the Nutanix Enterprise AI user interface, the management of API keys is located within the Endpoints section. To deactivate a key, the administrator must navigate to the Endpoints page from the left navigation pane.

Once on the Endpoints page, there is a specific button or tab labeledAPI Keys. Clicking this opens the API Keys list view. The administrator can then select the specific key (via checkbox) and use theActionsdropdown menu to selectDeactivate. This location is distinct from the "Settings" or "Admin Configuration" menus.

Comprehensive and Detailed Explanation From Exact Extract of Nutanix Artificial Intelligence documents knowledge without any external URL or links:

According to Nutanix Enterprise AI architecture recommendations for improving Retrieval-Augmented Generation (RAG) performance, when the issue is specifically related to the relevance of the retrieved documents (context) rather than the generation capability of the LLM, the standard solution is to introduce arerankingstep.

Problem Identification:The user explicitly states that when therightdocuments are used, the output is good, but 20% of the time the retrieval is generic or irrelevant. This indicates a precision issue in the retrieval phase, not a generation issue.

Solution (Reranking):Arerank model(or Cross-Encoder) is deployed as a post-retrieval step. It takes the top N documents retrieved by the initial embedding search and re-scores them based on their direct relevance to the specific query, re-ordering the list to ensure the most pertinent information is passed to the LLM. This improves precision without requiring the expensive and time-consuming process of replacing the embedding model or re-indexing the entire database.

Comprehensive and Detailed Explanation From Exact Extract of Nutanix Artificial Intelligence documents knowledge without any external URL or links:

In the Nutanix Enterprise AI "Models" dashboard, the Pending status has a specific definition related to resource availability.13 The documentation states: "Pending: The system is waiting for the resources required to save the LLM.14 For example, if the required storage space is not available, Nutanix Enterprise AI maintains a Pending status until the storage space becomes available."15

This indicates that the import process has been accepted but cannot proceed to the "Processing" (downloading) phase because the cluster lacks the necessary underlying resources (specifically storage capacity in the PVC/Files support) to ingest the model.

If the token were unauthorized, the status would typically result in aFailedstate.

If the model were custom/unvalidated, it would still process if resources were available.

Comprehensive and Detailed Explanation From Exact Extract of Nutanix Artificial Intelligence documents knowledge without any external URL or links:

Nutanix Enterprise AI exposes inference endpoints that are API-compatible with the OpenAI standard to ensure easy integration with existing AI ecosystem tools.4 Consequently, the correct endpoint path displayed in the NAI Dashboard and used for chat interaction is /api/v1/chat/completions.5

When an administrator views the "Endpoint Access" details in the dashboard to copy the curl command or Python snippet, the URL provided follows the standard OpenAI version 1 format: https:// /api/v1/chat/completions. This path ensures compatibility with libraries like LangChain or the official OpenAI Python SDK which default to the /v1/chat/completions structure.

Comprehensive and Detailed Explanation From Exact Extract of Nutanix Artificial Intelligence documents knowledge without any external URL or links:

When a model import completes but the model is not listed as available for deployment, the issue is almost always related tocompatibility or metadata validation.

Validation Check:NAI requires specific file formats (e.g., safetensors, pytorch_model.bin) and a valid configuration file (config.json) to recognize and serve a model. If the imported repository lacks these files, uses an unsupported architecture, or has a corrupted metadata file, the platform cannot register it as a valid "deployable" entity.

Action:The administrator must verify the source repository to ensure themodel format and metadataalign with NAI's supported standards (e.g., checking for the correct task type like "text-generation" in the config).

Other Options:

Public IPs (B) are network configurations unrelated to the internal model registry list.

API Keys (C) are for client-side access, not model management visibility.

Restarting the cluster (D) is a disruptive step that does not address the underlying data validation error.

Comprehensive and Detailed Explanation From Exact Extract of Nutanix Artificial Intelligence documents knowledge without any external URL or links:

For a secure deployment of Nutanix Enterprise AI, particularly regarding model storage and container registries, the platform enforces strict security protocols.

Disable HTTP (Option A):NAI requires secure connections. The registry should not expose insecure HTTP endpoints; it must be configured to serve traffic exclusively over HTTPS.

Trusted Certificate (Option C):While self-signed certificates can be technically made to work with extensive workaround configurations (adding CA to trust stores on all nodes), the "deployment prerequisites" and supported security stance for NAI state that the registry must be secured with atrusted TLS/SSL certificate(signed by a valid Certificate Authority). This ensures that the underlying Kubernetes cluster and NAI services can validate the registry's identity without connection errors or security warnings during image pulls.

Comprehensive and Detailed Explanation From Exact Extract of Nutanix Artificial Intelligence documents knowledge without any external URL or links:

According to the "Prerequisites" and "Storage Requirements" section of the Nutanix Enterprise AI installation guide, the administrator must provision a specific storage class to handle the shared access required for model files. The documentation explicitly states: "Ensure that you create a ReadWriteMany (RWX) storage class that enables static or dynamic provisioning of NFS shares for persistent volumes."2

This RWX mode is critical because it allows the persistent volume containing the large language models (LLMs) to be mounted as read-write by multiple nodes simultaneously. This ensures that as the inference service scales across different worker nodes, all pods can access the same underlying model data without conflict or duplication. RWO (ReadWriteOnce) would restrict the volume to a single node, which breaks the scalability of the AI cluster.

Comprehensive and Detailed Explanation From Exact Extract of Nutanix Artificial Intelligence documents knowledge without any external URL or links:

When deploying Nutanix Enterprise AI or its underlying Kubernetes infrastructure on public cloud managed services, specific configurations apply regarding hardware acceleration.

Azure AKS Specifics:When configuring an Azure Kubernetes Service (AKS) cluster to support GPU workloads for NAI, the underlying Virtual Machine (VM) images provided by Azure for GPU instances (such as the NC or NV series) come pre-configured with the necessary NVIDIA drivers.

Installation Procedure:Consequently, the documentation explicitly states that when setting up the GPU nodes or configuring the GPU Operator on AKS, the administrator mustskip the step to install the GPU driver. Attempting to reinstall the driver over the pre-existing Azure driver can cause conflicts, instability, or deployment failures. The setup should rely on the cloud provider's pre-installed driver stack.

Comprehensive and Detailed Explanation From Exact Extract of Nutanix Artificial Intelligence documents knowledge without any external URL or links:

When deploying Nutanix Enterprise AI (NAI) on the Nutanix Kubernetes Platform (NKP), the storage architecture requires specific protocols for different types of data.

Model Repository Storage:For the storage of Large Language Models (LLMs)—whether they are pre-validated models downloaded from public hubs or custom models imported manually—the platform requires aReadWriteMany (RWX)compatible storage class.

Protocol Selection:The documentation explicitly mandates the creation of anNFS(Network File System) export (specifically NFS v4 is often recommended) on a Nutanix Files server to fulfill this RWX requirement.

Deployment Workflow:During the infrastructure setup, the administrator must create a storage class that dynamically or statically provisions these NFS shares. This allows multiple pods (such as inference servers and model managers) to access the heavy model files simultaneously across different nodes in the Kubernetes cluster.

Why not others:

Blockstorage (ReadWriteOnce) is used for internal components like the database (nai-db) and logging, but it cannot support the shared access required for the model repository.

CIFS(SMB) is not the standard protocol used for Kubernetes RWX Persistent Volumes in this specific NAI deployment workflow.

Comprehensive and Detailed Explanation From Exact Extract of Nutanix Artificial Intelligence documents knowledge without any external URL or links:

In the Nutanix Enterprise AI "Endpoints" view, specific attributes provide details about the hardware backing the inference service.

Accelerator:This attribute explicitly displays thenumber and name of the GPUs per instance(e.g., "1x NVIDIA A10G"). It confirms the hardware acceleration resources assigned to the model.

Other Attributes:

Statusindicates the operational state (e.g., Active, Provisioning, Hibernated).

Inference Enginerefers to the software backend (e.g., vLLM, Python, Triton) rather than the physical GPU hardware.

Comprehensive and Detailed Explanation From Exact Extract of Nutanix Artificial Intelligence documents knowledge without any external URL or links:

According to the security prerequisites for deploying Nutanix Enterprise AI (especially in air-gapped or private environments), the container registry used to host the NAI images must be secured using HTTPS with trusted TLS/SSL certificates.

While Kubernetes can be configured to use insecure registries (HTTP or self-signed certs) via specific flag overrides (such as --insecure-registry), Nutanix Enterprise AI's supported and recommended configuration for enterprise deployments mandates a secure registry connection. This ensures that the container images (which may contain sensitive AI logic or model data) are transferred over an encrypted channel and that the identity of the registry is verified by the cluster nodes.