NVIDIA Dynamo Snapshot: Fast Startup for Inference Workloads on Kubernetes

Schwinn Saereesitthipitak, Dan Feigin, Vikram Sharma Mailthody and Maksim Khadkevich — May 2026

[Schwinn Saereesitthipitak](https://developer.nvidia.com/blog/author/schwinnsaereesitthipitak/), [Dan Feigin](https://developer.nvidia.com/blog/author/danfeigin/), [Vikram Sharma Mailthody](https://developer.nvidia.com/blog/author/vmailthody/) and [Maksim Khadkevich](https://developer.nvidia.com/blog/author/mkhadkevich/) — May 2026

Kubernetes checkpoint and restore lifecycle with NVIDIA Dynamo Snapshot.

Cold-starting inference replicas on Kubernetes can take minutes while engines load weights, warm kernels, and compile graphs. In our blog post, NVIDIA Dynamo Snapshot: Fast Startup for Inference Workloads on Kubernetes, we introduce Dynamo Snapshot, a checkpoint/restore approach that combines cuda-checkpoint, CRIU, and a privileged snapshot-agent DaemonSet to restore warm workers from shared storage. We also walk through KV cache unmapping, CRIU restore optimizations, and GPU Memory Service (GMS), which bring the gpt-oss-120b prototype below five seconds and reduce startup time by 21x.