It is tempting to describe a captured inference workload as one large snapshot. ColdSnap deliberately uses a more explicit boundary.
Four different assets
An activation can involve four assets with different compatibility and ownership rules:
- An OCI capsule for process, CUDA/NCCL, residual, and derived-cache state.
- An activation runtime owned by the matching engine adapter release.
- An optional native model payload containing content-addressed model bytes.
- The pinned original model snapshot used for recovery when an admitted native payload is unavailable.
The distinction prevents a driver-specific capsule from pretending to be a portable model distribution. It also allows byte-identical native payloads to be reused where their recorded contract permits.
Compatibility stays visible
The manager probes the placement and selects a snapshot driver supported across the participating hosts. ColdSnap then admits or rejects the requested combination using the artifact, provider, capability, and qualification records involved in that activation.
Unsupported pre-release formats fail admission. They are not silently upgraded during a restore.
Why OCI capsules
OCI gives the captured state a content-addressed distribution and publication boundary without requiring ColdSnap to own a bespoke artifact transport. The placement manager still decides where artifacts live, how credentials are supplied, and when material is staged onto targets.
For the detailed contract, see shared model payloads and driver residuals and the full architecture document.
Browse the ColdSnap documentation or inspect the source on GitHub.