Prerequisites

This article describes some prerequisites to consider before you start the installation process.

The PhariaAI installation process requires familiarity with Kubernetes and Helm charts.


Credentials

You need a user account with access to the Aleph Alpha JFrog artifactory. We will provide this to you.

On your deployment server

The documentation assumes you are using Linux or macOS for your installation, but this is not a requirement.
Aspect Requirements

Container orchestration platform

Kubernetes client v 1.29 and above. You can check this using kubectl version
Check your connectivity using kubectl get nodes

Package manager

Helm v 3.0 and above. You can check this using helm version

On your Kubernetes cluster

For hardware

Criteria Minimum requirements

GPU

  • Minimal Setup: 2 GPUs (with MIG), 3 GPUs (without MIG)

  • Recommended Setup: 6 GPUs (with MIG), 7 GPUs (without MIG)

    The actual number of GPUs depends on models selected for deployment.

  • Type: NVIDIA Ampere, Lovelace or Hopper generation. Currently, only NVIDIA GPUs are supported. Support for other vendors may be added in the future.

  • GPU Nodes: Your Kubernetes cluster must include GPU nodes to run the inference stack application pods.

  • During the finetuning of models, additional GPUs are required. See PhariaFinetuning service resource requirements.

CPU and memory

Object storage

  • Quantity: 3x

  • Type: minio or any other S3 backend type for PhariaData and PhariaFinetuning

  • Input and output operations (IOPS) maximum: 1000 or above

  • Throughput maximum: 100 Mb/s or above

Persistent volumes

  • Persistent volumes accessible by all GPU nodes in the cluster are essential for storing model weights

  • Ensure your persistent volumes are configured to be accessible across availability zones, if applicable in your environment

For software

Criteria Minimum requirements

Networking

Installed in a single namespace with open communication between all services in the namespace

NVIDIA GPU Operator

We strongly recommend using NVIDIA GPU Operator v.24 and above on default settings to manage NVIDIA drivers and libraries on your GPU nodes. More details on the GPU Operator setup can be found in NVIDIA GPU Operator setup.

Ingress controller and domain

  • The cluster must include an ingress controller to enable external access to the PhariaAI service

  • A certificate manager must also be configured to support secure access via Transport Layer Security (TLS)

  • A dedicated domain must be assigned to the Kubernetes cluster, enabling each service to host its application under a subdomain of this domain; for example, https://<service-name>.<ingress-domain>

Relational database management

Postgres v.14.0 or above

Quantity: 1x Large
Storage: 800 GB
CPU: 8x
Memory: 16 GB

Network access and whitelisting

Not required if networking requirements are met.
If you require multiple namespaces, contact our Product Support team.

Artifact management

Ability to pull the Helm chart containing the pharia-ai-helm and container images from an external artifact repository manager, such as JFrog. Credentials for this will be provided to you.

Monitoring and observability

No fixed requirements, but we recommend the use of Prometheus and Grafana

Cert manager

Cert Manager is required to provision webhook certs for the Dynamic Model Management feature

ClusterRole

PhariaOS requires a ClusterRole for hardware discovery and model management. By default, the chart creates the necessary ClusterRole and ClusterRoleBinding. For detailed configuration, see Using existing cluster roles for PhariaOS.