Review SeekrFlow requirements

SeekrFlow runs on a Kubernetes cluster in your own environment, on Amazon Web Services (AWS) or in your own data center. Before you begin, confirm you have everything on this page. Have your security team review Security considerations before you provision anything.

Kubernetes cluster

SeekrFlow supports conformant Kubernetes clusters running one of the three most recent minor versions, listed on the Kubernetes Releases page. Versions a cloud provider maintains beyond that, such as Amazon Elastic Kubernetes Service (Amazon EKS) extended support, aren’t supported. The cluster must run Metrics Server. Under heavy load, SeekrFlow adds copies of its services, and Kubernetes uses Metrics Server to measure that load. Without it, SeekrFlow can’t add copies. SeekrFlow works with policy engines such as Kyverno and Open Policy Agent (OPA) Gatekeeper. See Set up networking and cluster policies for the settings your policies need to permit.

Network connectivity

SeekrFlow doesn’t require a particular network design. It requires these connections:

The cluster to the container registry, object storage, PostgreSQL, OpenSearch, and any cloud model services you use.
Your users to the SeekrFlow load balancer.
The machine you install from to the Kubernetes API. If the API has no public endpoint, install from a machine inside your network, such as a bastion host.

For every outbound connection SeekrFlow makes, see Security considerations.

Worker nodes

Worker nodes run the platform services. For how many virtual CPUs (vCPUs) and how much memory they need, see Size your deployment.

Requirement	Detail
Nodes	At least 3 if you enable the Valkey caches. Each cache runs three pods on separate nodes.
Node type	Nodes you manage. Serverless compute, such as AWS Fargate, isn’t supported.
Root volume	At least 100 GiB.
CPU architecture	amd64 (x86-64). arm64, including AWS Graviton, isn’t supported.

GPU nodes

GPU nodes run SeekrFlow’s model servers, fine-tuning, and explainability workloads. Every deployment needs them. Amazon Bedrock can serve inference alongside them, but not fine-tuning or explainability.

Requirement	Detail
GPU vendor	NVIDIA. AMD and Intel GPUs are not supported.
CPU architecture	amd64 (x86-64). arm64 isn’t supported.
GPU architecture	NVIDIA Ampere or newer, meaning CUDA compute capability 8.0 or higher, such as A10, L4, L40S, A100, or H100. To check another GPU, see NVIDIA’s CUDA GPU compute capability list.
NVIDIA driver	Version 580 or newer. SeekrFlow’s container images include CUDA. You install only the driver on the nodes, not CUDA.
GPU memory per node	Enough for the largest model you run on that node. A model can’t span nodes.
Root volume	At least 200 GiB, mainly for the model server images. Fine-tuning jobs require additional space. See Size your deployment.

For how many GPUs your models need, see Size your deployment.

Services you provision

Service	What SeekrFlow uses it for	Set it up in
PostgreSQL 16 or later	Stores service data.	Provision PostgreSQL
OpenSearch	Stores the search indexes for file search and retrieval-augmented generation. AWS deployments provision it. On-premises deployments run it in the cluster.	Provision OpenSearch
Object storage	Stores datasets, model artifacts, and checkpoints.	Provision storage
Shared file storage	Caches model weights for GPU nodes.	Provision storage
Ingress	Routes your users to SeekrFlow, through an ingress controller, a DNS record, and a Transport Layer Security (TLS) certificate.	Set up networking and cluster policies
Secret store (optional)	Holds credentials outside the cluster.	Set up secret management

Credentials and access

Item	From	Purpose
License key	Your Seekr account team	Validated at startup. Services don’t run without it.
Registry credentials	Your Seekr account team	Pulls the Helm chart and container images.
Model configuration	Your Seekr account team	Defines which models to run on your GPU nodes and how to size them.
Hugging Face access token	Your Hugging Face account	Downloads gated model weights.
Brave Search API key	The Brave Search API	Lets agents run web searches.
Access to Amazon Bedrock models (optional)	Your cloud account	Serves inference through externally hosted models.
Kubernetes cluster-admin access	Your cluster administrator	Installs the chart’s cluster-wide resources, such as custom resource definitions and roles. SeekrFlow has no reduced-permission install.

For how to get each one, see Request credentials.

Tools

The install runs from a laptop or bastion host that can reach the cluster.

Tool	Used for
`sfe` command-line interface (CLI)	The recommended install path. Download it from Seekr’s registry for Linux on amd64 or macOS on Apple silicon. See Set up the sfe CLI.
`kubectl`	Inspecting the cluster, and the steps you run without `sfe`.
`helm`	Installing prerequisites such as the AWS Load Balancer Controller and External Secrets Operator. For the Helm install path, use Helm 3.2.0 or later.
AWS Command Line Interface (AWS CLI)	Signing in to your AWS account, which `sfe` uses through the AWS SDK, and setting up resources without `sfe`.
Docker	Signing in to registries before you mirror container images. Needed only if you mirror.

SeekrFlow Self-Managed

Plan a SeekrFlow deployment

Provision SeekrFlow dependencies

Prepare to install SeekrFlow

Deploy SeekrFlow

Operate SeekrFlow

Reference

Kubernetes cluster

Network connectivity

Worker nodes

GPU nodes

Services you provision

Credentials and access

Tools

​Kubernetes cluster

​Network connectivity

​Worker nodes

​GPU nodes

​Services you provision

​Credentials and access

​Tools

Kubernetes cluster

Network connectivity

Worker nodes

GPU nodes

Services you provision

Credentials and access

Tools