Nvidia Dynamo Snapshot Cuts Inference Startup to Seconds on Kubernetes
Nvidia’s Dynamo Snapshot combines CRIU and CUDA checkpointing to checkpoint and restore GPU‑accelerated inference containers on Kubernetes, shrinking cold‑start times from minutes to seconds. The prototype restores a full vLLM worker on a new node with near‑instant readiness, helping avoid SLA breaches during traffic spikes. The approach is detailed in an official Nvidia developer blog.