dstack vs KAI Scheduler

dstack is an open-source orchestration layer for AI workloads across GPU clouds, Kubernetes, virtual machines, and bare-metal clusters. YAML configurations define fleets, development environments, tasks, services, presets, and volumes, while dstack handles provisioning and job scheduling, including auto-scaling, port forwarding, and ingress. Supported accelerators include NVIDIA, AMD, TPU, and Tenstorrent. Documented backends include AWS, Azure, GCP, Kubernetes, GPU cloud providers, remote SSH hosts, and experimental Slurm. Tasks can use frameworks such as accelerate, torchrun, Ray, and Spark. Users manage resources through the CLI or HTTP API, and the server can run on a laptop or another environment that can access the clusters in use. Services can publish inference endpoints through gateways with HTTPS, custom domains, auto-scaling, and rate limits. The self-hosted open-source stack is free. Server data and project secrets are stored in plaintext by default unless administrators configure AES-256-GCM encryption. TPU support is limited to single-host instances with up to eight cores. GPU Marketplace usage is prepaid, with prices that vary by GPU and provider and appear in the console before provisioning.

KAI Scheduler is a self-hosted, open-source Kubernetes scheduler for allocating GPUs to AI workloads such as data processing, training, and inference. It supports interactive notebooks through multi-node distributed training in one cluster and is designed for large GPU clusters and high-throughput workloads. Hierarchical queue trees provide quotas, limits, priorities, and borrowing between teams. Gang scheduling places distributed workloads together when resources are available, while elastic workloads change replica counts within configured minimum and maximum values. GPU sharing includes time slicing, MPS, and MIG. Other scheduling controls include topology-aware placement, time-based fairshare, and priority-based preemption. KAI documents integrations with Ray on Kubernetes and Grove. It runs on cloud or on-premises infrastructure, including Karpenter environments. Installation uses a production Helm chart, source builds, or ArgoCD, with a running Kubernetes cluster, Helm CLI, and NVIDIA GPU Operator required for GPU requests. The Apache-2.0 project is also a CNCF sandbox project.