# dstack A control plane for provisioning and orchestrating GPU workloads across cloud, Kubernetes, and on-prem infrastructure. - Findings24: - Site: - Preços: Freemium - Categorias: Plataformas de inferência de IA, Implantação de modelos de IA, Nuvem de GPU ## O que é dstack? dstack is a unified control plane for provisioning and orchestrating GPU compute across GPU cloud providers, Kubernetes, and on-premises clusters. It coordinates infrastructure with AI development, training, and inference workflows, and supports NVIDIA, AMD, Google Cloud TPU, and Tenstorrent accelerators. Users describe fleets and workloads in YAML configuration files stored with their project, then apply them through the `dstack` CLI or programmatic API. dstack handles infrastructure provisioning and job scheduling, with related capabilities including autoscaling, port forwarding, and ingress. The documented workload types include remote development environments, tasks for training or batch processing, and services that expose models or web applications as endpoints. ## O que dstack pode fazer? ### Infrastructure across environments Configure GPU cloud backends, Kubernetes clusters, or SSH-accessible on-premises hosts, and manage compute through a common control plane. ### Declarative workload configurations Define fleets, development environments, tasks, and services in YAML files, then apply them with the CLI or programmatic API. ### Provisioning and scheduling dstack provisions infrastructure and schedules submitted workloads. Fleets can be configured with node and GPU ranges, including templates that defer instance provisioning until runs are submitted. ### Multi-node workloads Tasks can run across clusters by specifying nodes or node groups; fleet placement can also be configured for cluster use. ### Endpoints for services Deploy models or web applications as services. Service replicas can be configured behind a single endpoint. ### Multiple accelerator families The documentation lists out-of-the-box support for NVIDIA, AMD, Google Cloud TPU, and Tenstorrent accelerators. ## Casos de uso ### Develop remotely with an IDE Create a development environment on provisioned compute and connect using a desktop IDE such as VS Code or through SSH. ### Run training and batch jobs Submit command-based tasks for model training, batch processing, or other jobs, with configurations for single-node or multi-node execution. ### Serve a model or web application Package a service configuration with its command, port, model identifier where applicable, and resource requirements to deploy an endpoint. ### Manage a shared on-prem compute pool Register SSH-accessible servers as a fleet so workloads can use existing hosts that meet dstack's stated host requirements. ## Perguntas frequentes ### What infrastructure can dstack manage? The documentation describes GPU cloud providers, Kubernetes, and on-premises clusters. On-prem hosts can be configured through SSH fleets, subject to the documented host requirements. ### How are workloads submitted? Define a workload configuration, commonly as a YAML file, and apply it with the \`dstack apply\` CLI command. The documentation also describes a programmatic API. ### What workload types are supported? The documented configuration types include development environments, tasks, and services. Fleets define and manage pools of compute instances. ### Can tasks use multiple nodes? Yes. The quickstart says tasks can run on a cluster by specifying \`nodes\` or defining multiple node groups. ### Does dstack include a hosted option? The documentation says users can install the dstack server or sign up for dstack Sky. ## Informações rápidas - **Category**: GPU infrastructure orchestration - **Platform**: GPU clouds, Kubernetes, and on-premises clusters - **Accelerators**: NVIDIA, AMD, Google Cloud TPU, and Tenstorrent - **Workload configurations**: Fleets, development environments, tasks, services, and experimental inference presets - **Configuration workflow**: YAML files applied with the dstack CLI or programmatic API - **Official domain**: dstack.ai