Chapter 1. Storing models in disconnected environments
You can store language models in disconnected environments using OCI-compliant model container images or persistent storage volumes. Each approach has different tradeoffs for deployment complexity, storage efficiency, and operational workflows.
Use OCI model images when you want to use Red Hat validated models and prefer a unified container-based workflow.
Use persistent storage when you need to deploy custom models, fine-tuned models, or models not available as OCI images.
- OCI model container images
OCI-compliant model container images, also known as modelcars, package language models as container images that you can store in your mirror registry alongside other container images. This approach integrates with existing container image workflows and infrastructure:
- Uses the same mirroring workflow as other container images
- Leverages existing container registry infrastructure for versioning and distribution
- Enables faster pod startup through image caching on nodes
- Simplifies model updates through standard image pull mechanisms
NoteOCI model container images require additional registry storage capacity for large model images. Model images can be 10-100 GB depending on model size and applied quantization.
Red Hat provides validated OCI model images in the
registry.redhat.io/rhelai1namespace that you can mirror to your disconnected registry.- Persistent storage volumes
You can store model files directly with persistent storage such as Network File System (NFS) volumes or other persistent volume types supported by OpenShift Container Platform. This approach requires transferring model files to the disconnected environment separately from container images. You can share a single copy of a model across multiple inference pods with the same persistent storage volume. You can store models downloaded from Hugging Face or other sources, or you can store custom or fine-tuned models not available as OCI images.
Persistent storage volumes require separate transfer and setup workflow for model files, with appropriate storage provisioning and access configuration.