NVIDIA has released Snapshot to checkpoint GPU workloads after their initial preparation and restore new copies from that state. The Kubernetes tool preserves processor and GPU memory so replicas can avoid repeating model loading and compilation. Restoration requires compatible hardware and drivers; container downloads, infrastructure provisioning and active conversation caches remain outside the saved startup state.
NVIDIA released Snapshot to save GPU workloads after their initialization and start additional copies from the saved state. The open-source tool runs on Kubernetes, the system used to manage containerised applications. It captures memory from both the central processor and the graphics processing unit, or GPU. A new inference replica can then resume without repeating model-weight loading, compilation and preparation of CUDA execution graphs.[1], [2]
The destination node must use the same GPU model and driver as the saved workload. Compatible jobs spread across several GPUs on a single node can be captured together. Kubernetes resources can checkpoint an existing running pod or launch a workload and save it once ready. The application also needs a small handshake: it pauses before capture and resumes after restoration. Checkpoint creation and storage have their own upfront costs.[1]
Snapshot reuses the preparation already completed inside a workload. Downloading a container image, preparing its runtime and provisioning infrastructure still occur outside the restore path described by NVIDIA. The key-value cache used for ongoing model context is released before checkpointing and excluded from the artifact. These boundaries distinguish reusable initialization from automatic transfer of active conversations. The system’s documented scope includes compatible workloads within a single node, rather than arbitrary movement between different machines and hardware.[1]
