2 Requirements #
Deploying SUSE AI Factory in an air-gapped environment requires specific local infrastructure components. You must prepare a connected staging environment to download resources. You also need to configure OCI-compliant local registries to host Helm charts and container images. Secure TLS endpoints protect internal network communication.
2.1 Connected environment with adequate storage #
Mirroring OCI artifacts and models requires a connected environment with sufficient storage. Use this connected environment to download all required files.
2.2 OCI-compliant artifact registries #
You must set up one or more OCI-compliant artifact registries inside the air-gapped environment. These registries host mirrored artifacts such as Helm charts and container images. Configure these registries with production TLS certificates to secure the connection endpoints.
When mirroring artifacts, apply the following rules:
Use one-to-one mirroring. Mirror all artifacts from a specific source registry into the same destination registry. You can mirror multiple source registries into a single destination registry if you use unique subpaths or namespaces to distinguish them.
Preserve repository paths. For example, mirror the repository
foo/bar:1.0.0fromoci://acme.com/foo/bar:1.0.0to your local registryoci://mylocal.registry.com. You can place the mirrored repository under a subpath, such asoci://mylocal.registry.com/mysubpath/foo/bar:1.0.0. However, you must preserve the original repository name and tag,foo/bar:1.0.0.
2.3 Distributed storage provisioner #
Persistent Volume Claims (PVCs) facilitate the model cache. For multi-node cluster deployments, the storage provisioner must support distributed storage, such as Longhorn or NFS. These PVCs require ReadWriteMany (RWX) access.
2.4 Registry mirroring tooling #
To mirror OCI artifacts, you require mirroring tools such as regsync, Skopeo or crane. These tools pull artifacts from source registries and push them to destination registries. You can also use artifact-specific tools, such as Helm to mirror Helm charts, and Docker or Podman to mirror container images.
2.5 Container runtime environment #
Caching NVIDIA NIM models from the NGC catalog requires NIM model puller containers. You must run the model puller container in a connected environment using Docker or Podman. This environment downloads and caches the model data locally.
2.6 Hugging Face CLI (optional) #
If you require Hugging Face models, use the Hugging Face CLI to download the models to the local cache.