Run published capsules
Materialize locally → run. Prepare a published recipe for your hosts, then launch it with sparkrun run.
ColdSnap documentation
ColdSnap brings compatible vLLM and SGLang workloads online from captured runtime state. Start with Sparkrun to run community capsules or create your own.
Use ColdSnap with Sparkrun
Section titled “Use ColdSnap with Sparkrun”First, set up the ColdSnap plugin preview and your Sparkrun cluster. The plugin currently requires this development environment; bundled Sparkrun support is coming soon. Then choose your workflow:
Capture and share your work
Capture → run and measure → publish. Create a local capture, check its responses and performance, and publish when you are satisfied.
Use sleep, wake, and warm to release GPU memory between uses or prepare a workload ahead of demand. The benchmark matrix compares default loading, InstantTensor, ColdSnap recovery, and ColdSnap native.
For detailed Sparkrun configuration, continue to the Sparkrun reference, recipe policy and weights, or plugin internals. These live under Community · Sparkrun → Advanced usage.
Develop and integrate ColdSnap
Section titled “Develop and integrate ColdSnap”The advanced developer docs explain ColdSnap itself: the controller, engine integrations, snapshot drivers, artifacts, and manager contracts. Start there when building an integration or working on ColdSnap internals.
Sparkrun is the reference manager. Other managers can implement the same contracts; Docker is the reference runtime backend, and Kubernetes is not yet implemented or qualified.