Skip to content

ColdSnap for advanced developers

These docs cover ColdSnap itself: the Go controller, engine adapters, Python/native integrations, artifacts, and manager contracts. They are for developers extending ColdSnap or implementing an integration.

To run community capsules with Sparkrun, start with the community guides.

Start with architecture, then execution topology and model payloads and residuals. Captures separate driver-qualified runtime state from large model weights.

The n580 driver restores a pre-CUDA process template and reconstructs fresh CUDA state. The n610 driver restores initialized CUDA process state and preserves qualified graph executables by default. Compatibility remains specific to the engine, model, driver, runtime, hardware, and topology.

The direct coldsnap commands take strict JSON requests and require a manager host-provider. They do not discover or connect to remote hosts themselves.

Recipe configuration, CLI workflows, and plugin implementation details are in Community · Sparkrun advanced usage.

Read the operator integration contract, host-provider protocol, and runtime-neutral manager contract together.

ColdSnap 0.3.20 requires runtime-v1. The manager owns placement, transport, credentials, image/workload execution, and workload identity. Docker is the reference backend. Kubernetes is not yet implemented or qualified.

For a new engine, use the vLLM and SGLang guides as reference implementations. The engine adapter consumes the shared request and artifact contracts; its runtime integration owns model loading, memory regions, graph and communication lifecycle, and response validation. Keep scheduling, transport, and credentials in the manager.

Technical reference pages are synchronized from the ColdSnap source repository; their edit links point to the canonical documents. The benchmark hub describes the current measured scope separately from these implementation contracts.