Start here

ColdSnap captures a fully initialized inference workload so later activations can restore validated runtime state instead of repeating model construction, compilation, calibration, distributed setup, and CUDA graph capture.

Before you begin

ColdSnap is designed to be driven by a placement-aware manager. Sparkrun is the reference manager today and is the recommended user-facing path. The lower-level coldsnap commands expose strict manager-facing JSON contracts; they are not a replacement scheduler.

You will need:

• A supported Linux and NVIDIA driver combination.
• Digest-pinned vLLM or SGLang runtime images containing the matching integration.
• A pinned model revision and access to its model payloads.
• A Sparkrun recipe and cluster placement, or your own implementation of the operator contract.

Qualification is exact

Support is determined by the complete engine, model, runtime image, CUDA, snapshot driver, hardware, and topology combination—not by any one version number alone.

Recommended reading order

1. Run the quickstart.
2. Read the full capture, publish, and restore guide.
3. Check the configuration reference.
4. Review recovery-aware operations before production use.
5. Use the architecture document when integrating a new manager or engine.

Key terms

Capsule : A driver-qualified OCI artifact containing process, CUDA/NCCL, residual, and derived-cache state for a launch unit.

Activation runtime : The small, adapter-owned controller staged onto each target host and mounted read-only into rank containers.

Native model payload : Optional content-addressed model bytes captured in the engine’s native representation and reusable where compatibility permits.

Recovery : Reconstructing model state from the pinned original safetensors when no admitted native payload is used.

Qualification : Recorded evidence that a specific configuration restores and serves the expected validation response within its declared policy.