Licensed to be used in conjunction with basebox, only.
// installation
Requirements
Applies to
Product: Server · Audience: Platform Operator
What you need before you start: hardware (see reference configurations), operating system, network access to registries, DNS and TLS decisions, model choice. Settle these points before you touch the server – most delays in installations come from open decisions, not from technical problems.
Hardware
| Component | Minimum | Recommendation |
|---|---|---|
| GPU | NVIDIA with compute capability 7.0+ (V100, T4, RTX 20xx, A100, L4, H100, H200 …) | Per reference configuration; MIG-capable if service models are to run via MIG |
| CPU | 8 cores | 16+ cores |
| RAM | 16 GB | 32 GB+, 64 GB+ for large models; FAST LTA systems: 128 or 512 GB |
| Storage | 500 GB | 1 TB+ SSD/NVMe |
| Power, cooling | For all GPUs at their power limit | Approved by the manufacturer |
Check after the OS installation: lspci | grep -i nvidia, lscpu, free -h, df -h.
What decides the GPU question is not basebox but inference and service models – see Hardware options. A CPU-only application server with a separate GPU host is also possible: Deployment topologies.
Operating system
Ubuntu 24.04 LTS Server (x86_64) for new installations; 22.04 LTS for existing systems. Debian, RHEL, SLES, Rocky/Alma may work but are not officially tested; Windows Server is not supported (vLLM requires Linux). Details and reasoning: Server Preparation Guide → Alternative operating systems.
Network access
During installation the server needs outbound HTTPS (443) to package sources and registries – Ubuntu, Docker, Kubernetes (pkgs.k8s.io), NVIDIA (developer.download.nvidia.com, nvidia.github.io, helm.ngc.nvidia.com), Helm, container registries, Calico – and to the basebox registry gitea.basebox.health. The full domain list with test commands: Required network access.
Inbound: 22/TCP (SSH), 6443/TCP (Kubernetes API, if external), 80/443 (ingress).
Air-gapped environments are supported – via an internal image mirror and pre-loaded model artefacts. Agree the procedure in advance with support@basebox.ai.
Registry access
The umbrella chart lives at oci://gitea.basebox.health/basebox-distribution/helm/basebox.ai, the images under gitea.basebox.health/basebox-distribution/. Check pull access from the nodes' container runtime, not just from your workstation. Credentials for the OCI login are under Using Helm charts.
Decisions to make beforehand
| Decision | Options | Where to continue |
|---|---|---|
| Domain and DNS | Public domain · internal domain · basebox.local (evaluation only) |
Networking |
| TLS mode | cert-manager (public domain) · existing-secret (own certificate, internal CA) · local (evaluation) |
Networking |
| Language model | Per GPU equipment; context size and concurrency | Models & inference · LLM recommendations |
| Inference | Bundled vLLM on the same node · external endpoint | Connect inference |
| Service models | Dedicated GPU · MIG slices · CPU mode | Deploy service models |
| Identity | Local accounts · LDAP/AD · OIDC SSO | LDAP / OIDC |
| SMTP server for invitations · invitation links manually at first | SMTP | |
| Connectors | Which ones; endpoints per connector | MCP connectors with Helm |
| Operation | Self-managed · by basebox | Operating models |
Access and tools
- Physical or remote access to the server (IPMI, iLO or similar), root privileges, SSH.
- On your workstation:
kubectl,helm(3.x),curl;rgorgrepfor image checks. - A place for credentials: password manager or secret store. The installation generates admin passwords that you have to hand over securely.
- If basebox is commissioned to install or operate: approved PAM/VPN access – see Remote maintenance.
Checklist
- Hardware matches a reference configuration or is agreed with basebox as Custom
- GPUs recognised by the system (
lspci | grep -i nvidia) - Ubuntu 24.04 LTS Server ISO and installation medium ready
- Network access to all domains checked – or offline transfer agreed
- Registry pull tested from the runtime
- Domain, DNS and TLS mode decided; certificate or ClusterIssuer ready
- Language model chosen and checked against VRAM
- Service models: GPU allocation or CPU mode decided
- Identity and mail integration planned
- Backup target for databases planned (S3-compatible storage for CloudNativePG or dump location)
Next step: Server Preparation Guide, then NVIDIA / GPU