Overview
The purpose of the AI Factory blueprint collection is to provide a step-by-step process for setting up a simple AI Factory system and getting it up and running quickly, including:
Identifying the minimum hardware and networking requirements for your AI Factory. These baseline specifications also serve as a reference for more advanced deployments. OpenNebula supports high-performance architectures such as InfiniBand, Spectrum-X, and NVLink, although these setups are not automated and require custom configuration.
Follow the step-by-step deployment instructions using OneDeploy to build your AI Factory, with options for both on-premises installations and cloud-based deployments.
Optionally, you can validate the setup using the same methodology we apply during formal infrastructure acceptance. This validation covers direct vLLM execution for inference, SLURM integration for fine-tuning, and Kubernetes-based execution using NVIDIA Dynamo® for inference and NVIDIA KAI Scheduler® .
Hardware Requirements
AI Factories require high performance server hardware, including accelerated infrastructure such as GPUs and high-performance networking. The following details outline a minimal recommended hardware configuration:
- CPU: 16 physical or logical CPU cores with hardware virtualization support.
- Memory: 64 GB RAM. More memory is recommended when running multiple Kubernetes workers or large AI models.
- Storage: At least 250 GB of fast SSD or NVMe storage, with sufficient additional capacity for VM images, container images, model weights, and datasets.
- GPU: At least one NVIDIA data-center GPU with PCI passthrough support. Two GPUs are recommended for testing disaggregated inference workloads such as separate prefill and decode workers. See the Platform Notes for supported GPU models.
- Networking: At least one physical Ethernet interface suitable for bridged OpenNebula networking. A 10 GbE or faster network is recommended for multi-node deployments.
- Virtualization: CPU virtualization extensions and IOMMU enabled in the system firmware to support KVM and PCI passthrough.
- Operating system: Ubuntu 24.04 is highly recommended. See the Platform Notes for a list of alternative operating systems supported by OpenNebula.
- Bare-metal Cloud Instances: If you are using resources provided by a 3rd-party cloud provider, bare-metal instances are highly recommended. AI Factory deployments on virtual instances will result in heavily degraded performance and may not function at all.
Basic Outline
Configuring, deploying and validating a high-performance AI infrastructure using OpenNebula involves these steps:
Familiarize yourself with Architecture and Specifications. We recommend consulting the guide on GPU PCI-passthrough for details relating to your GPU hardware and IOMMU.
Deploy and configure your AI Factory with one of these alternatives:
- On-premises AI Factory Deployment: Set up an AI Factory using OneDeploy for On-premise environments.
- On-cloud AI Factory Deployment: Set up an AI Factory using OneDeploy with Scaleway for cloud environments.
Integrate external infrastructure services if required:
- NVIDIA InfraControler (NICo): Offer multi-tenant bare metal instances from an existing OpenNebula cloud.
- NVIDIA InfraControler (NICo): Offer multi-tenant bare metal instances from an existing OpenNebula cloud.
Perform Validation: As a prerequisite, you must have an AI Factory ready to be validated after completing the above installation procedures. These are the options to validate your AI Factory:
- LLM Inferencing with vLLM: Using vLLM with two different models and two model sizes, running across both H100 and L40S GPUs.
- LLM Fine-Tuning with NVIDIA Slurm: Fine tuning an AI model using the OpenNebula NVIDIA Slurm appliance.
- Deployment of AI-Ready Kubernetes: Use H100 and L40S deployment to run Kubernetes.
- LLM Inferencing with NVIDIA Dynamo: Integrating the GPU-powered Kubernetes Cluster with the NVIDIA Dynamo Cloud Platform to provision and manage AI workloads through the Dynamo framework for your AI workloads on top of the NVIDIA Dynamo framework.
- Scheduling with NVIDIA KAI Scheduler: Use the NVIDIA KAI Scheduler to share GPU resources across different workloads within the AI-ready Kubernetes Cluster.
GIVE FEEDBACK
Was this resource helpful?
Glad to hear it
Sorry to hear that