Inference servers do valuable work before they can serve a first token. They construct the model, establish the distributed topology, compile runtime kernels, calibrate memory, and capture CUDA graphs. Conventional starts repeat that work whenever a workload comes online.
ColdSnap asks a narrower question: when the workload and its operating environment are compatible, how much of that initialized state can be treated as an infrastructure artifact?
Pay once, validate every time
ColdSnap starts the real vLLM or SGLang workload, validates it, and captures its admitted process and GPU runtime state. A later activation restores from that state, hydrates model data using the selected weight provider, and validates the resulting inference response before returning control to the manager.
That last validation matters. The goal is not merely to reconstruct processes; it is to return a working serving workload with an explicit receipt.
A manager-driven boundary
ColdSnap is not a scheduler or another cluster daemon. A placement-aware manager owns hardware discovery, credentials, image and model staging, transport, and workload identity. ColdSnap owns strict requests, compatibility admission, engine coordination, artifacts, lifecycle transitions, and receipts.
Sparkrun is the reference manager today. The same boundary is intended to support other managers without embedding their placement or transport policy inside ColdSnap.
Artifacts, not duplicated model repositories
Captures separate large, portable model data from driver- and runtime-qualified state. OCI capsules carry the state that belongs to the snapshot boundary, while native payloads or pinned source safetensors provide model bytes. This makes compatibility explicit and avoids hiding a second model registry inside the system.
Open development
ColdSnap is pre-release research software released under AGPL-3.0-only. Its support claims remain limited to configurations with recorded qualification evidence.
Read the documentation, inspect the architecture, or review the accepted benchmarks.
Browse the ColdSnap documentation or inspect the source on GitHub.