Skip to content
Copyright © 2022-2026 basebox GmbH, all rights reserved.
Licensed to be used in conjunction with basebox, only.

// installation

Bare-metal installation

Applies to

Product: Server · Audience: Platform Operator

The end-to-end path for self-installers: from an empty server to a validated basebox installation. Follow the steps in this order; each page covers exactly one step, refers to the authoritative guides and ends with a check before you move on. If you already have a suitable Kubernetes cluster, join at step 7.

The steps

# Step Result Page
1 Requirements Hardware, network, decisions on DNS/TLS/model settled Requirements
2 Prepare the server Ubuntu 24.04 LTS installed, updated, network access checked Server Preparation Guide
3 NVIDIA / GPU Driver, CUDA, container runtime, GPU Operator; MIG where intended; node advertises GPU resources NVIDIA / GPU
4 Kubernetes Cluster with ingress controller, storage class, CloudNativePG Kubernetes
5 Storage Volumes for databases, media, model artefacts planned Storage
6 Networking DNS, TLS mode, firewall, boundary AISRV ↔ inference Networking
7 Install basebox Umbrella chart via Helm, pods running, credentials retrieved Install basebox
8 Deploy service models ragsrv, ragsrv-support, OCR, STT on GPU or CPU, healthy Deploy service models
9 Connect inference Bundled vLLM or external endpoint, model responds Connect inference
10 Validate installation Acceptance: chat, RAG, OCR, STT, reboot, manifest recorded Validate installation

Time required

With prepared hardware and network access: steps 2–4 a few hours (including model downloads), step 7 under an hour, steps 8–10 one to two hours. Allow one working day in total; air-gapped environments additionally need the offline transfer of images and models.

What you should bring

  • Linux and Kubernetes fundamentals; experience with kubectl and helm.
  • Access to the server (physical or IPMI/iLO), root privileges.
  • Network access to package sources, NVIDIA, Kubernetes repositories and the basebox registry gitea.basebox.health – or an offline transfer procedure.
  • A decision on which language model should run (Models & inference) and whether service models run on GPU or CPU.

What this path assumes

  • One server with NVIDIA GPU(s) per a reference configuration – or agreed with basebox as Custom. For separate application and inference hosts, the path applies to both hosts; the application server then needs no NVIDIA stack, see Deployment topologies.
  • Ubuntu 24.04 LTS. Other distributions may work but are not officially tested.
  • Kubernetes 1.33+ recommended (the chart requires 1.23+).

Verified state

This path was run in February 2026 with Ubuntu 24.04.3 LTS, NVIDIA driver 580.126.09, CUDA 13.0.88, Docker 29.2.1, Kubernetes 1.33.7, Helm 3.20.0 and GPU Operator (latest). Versions change quickly; every step page links the official documentation, which remains authoritative.

Principles for the whole path

  1. Check first, then continue. Every step ends with a command that confirms the state. Do not move on while it does not match.
  2. Assign GPUs by stable identity (nvidia-smi -L: UUID, PCI address), not by index numbers, which can change after reboots.
  3. Never put secrets in values files or Git. Database passwords, API keys, registry and inference credentials belong in Kubernetes secrets.
  4. Record versions. Chart, app, driver, CUDA, Kubernetes, GPU Operator, image digests – you need the manifest for every update and every support case.

Help

Next step: Requirements