Cedana

Cedana is infrastructure for checkpointing, migrating and resuming running workloads, with a focus on GPU inference and training. It saves the full state of a live job, including CUDA context, GPU memory with model weights and KV cache, and open network connections, moves it to another node over RDMA, and restores it where it left off. The aim is to stop compute being stranded by idle GPUs, failures, overprovisioning and rigid scheduling. It is for teams operating GPU fleets, from AI labs and inference providers to HPC groups. Cedana presents one API across CPU and GPU containers, distributed jobs, Kubernetes pods and SLURM batch, and works transparently with MPI and NCCL. Use cases include surviving node loss, preemption and hardware faults without restarting, running on spot capacity and riding out reclaims, consolidating utilisation, performing maintenance without losing progress, and resizing across instances, clusters, regions and clouds. It deploys as a Kubernetes Helm chart or a SLURM plugin, requires no code changes, and lists AWS, Google Cloud, NVIDIA, Kubernetes, SLURM and NVIDIA Dynamo among its integrations. The site also demonstrates warm restores of large models as a faster alternative to cold starts. Pricing is not published. Access is through a scheduled demo or a request for access, so costs are quote based. The company describes its longer term goal as building an operating system for AI factories, positioning migration as the base for policy driven orchestration that moves jobs in response to failures, demand, priorities or maintenance.