What is dstack?
A simplified substitute for Slurm and Kubernetes, dstacks was created especially for artificial intelligence. It speeds up the creation, training, and deployment of AI models by making container orchestration for AI workloads in the cloud and on-premises simpler.
dstack is simple to use with both on-premises servers and any cloud provider.
Container orchestration using AI teams
Dstack is an open-source substitute for Slurm and Kubernetes that makes container orchestration for AI workloads easier for both on-premises and cloud environments. Accelerators from NVIDIA, AMD, TPU, and Intel are supported.
Development environments
With only one command, you may supply a remote workstation with your code and preferred IDE in development environments.
Before scheduling a task or launching a service, you can use your preferred IDE or notebook to run code interactively in development environments.
Activities
You can execute a web application or schedule a job using a task. It enables you to set up ports, resources, dependencies, and other things. Clusters can be used to distribute and execute tasks.
Tasks are perfect for running apps for development or for training and optimising jobs.
Services
Deploying web apps or models as private or public auto-scalable endpoints is made possible by services. Auto-scaling rules, resources, dependencies, and authorisation can all be set up.
Any team member can utilise the web application or model after it has been deployed.
Fleets
Fleets make it possible to provision and manage clusters and instances both on-premises and in the cloud with efficiency.
Dev environments, jobs, and services can all reuse a fleet once it has been built.
Set up the server
Make sure you have installed the server or registered for dstack Sky before using dstacks.
Describe the configurations
The following setups are supported by dstack:
- Developer environments for desktop IDE-based interactive development
- Jobs that need to be scheduled, including distributed ones (or web apps)
- Model deployment services (or web apps)
- Fleets for on-premises and cloud cluster management
- Network volume management volumes (to persist data)
- YAML files in your repository can be used to define gateways for publishing services with a custom domain and HTTPS configuration.
Utilize configurations
Use a programmatic API or the dstacks apply CLI command to apply the configuration.
dstack handles auto-scaling, port-forwarding, ingress, and other tasks automatically, including job scheduling and infrastructure provisioning.
Why dstacks?
The creator and CEO of dstack describe the difficulties that dstack faces for teams working on AI and operations.
AI teams can work with any framework on on-premise servers and cloud platforms because to dstacks ability to simplify infrastructure administration and container usage.
How does it compare to other tools?
Comparing dstack to Kubernetes
With dstack’s increased portability and AI-specific design, AI engineers can manage development, training, and deployment without requiring additional tools or operations support.
With dstacks, everything is available right out of the box, so you don’t require Kubeflow or other Machine Learning platforms.
Furthermore, dstack is considerably simpler to use with on-premise servers; all you need to do is supply hostnames and SSH passwords, and dstack will automatically generate a fleet that is prepared for usage with tasks, services, and development environments.
Using Kubernetes and dstack together
Using dstacks directly with your cloud accounts or on-premise servers without Kubernetes is more efficient for AI development.
You can, however, configure the dstack server with a Kubernetes backend to provision via Kubernetes if you’d like.
Is Kubernetes the preferred choice for production-grade deployment by your operations team? Kubernetes and dstacks can be used in tandem; Kubernetes is for production-grade deployment, while dstack is for development.
Comparing dstack to KubeFlow
Kubeflow can be completely replaced with dstack. It includes all of Kubeflow’s functionality as well as a lot more, like development environments, services, and more features.
Because it doesn’t require Kubernetes and is compatible with a variety of cloud providers by default, dstack is simpler to set up with on-premises servers.
dstack versus Slurm
Slurm can be completely replaced by dstacks. It includes all of Slurm’s features plus a lot more, like services, development environments, cloud support that goes beyond the box, simpler setup with on-premises servers, and much more.
dstack: Now supporting Intel Tiber AI Cloud and Intel Gaudi AI Accelerators
Whether executing workloads in the cloud or on-premises, AI teams require flexibility. As a result, dstack, a participant in Intel’s Liftoff program for AI startups, is now integrating with Intel Tiber AI Cloud and supporting Intel Gaudi AI Accelerators.
This most recent version of dstack makes it simpler to manage and scale AI workloads across various environments by enabling the orchestration of containers across on-premises computers with Intel Gaudi accelerators.
How dstack Uses Intel Gaudi to Improve AI Deployment
Dstack makes it easier to manage AI infrastructure in both on-premises and cloud environments by integrating Intel Gaudi. Using Intel Gaudi’s high-performance architecture, developers can easily manage, deploy, and fine-tune AI models.
Gaudi-powered devices can be integrated with dstacks unified AI workflow. Businesses may drastically lower training and inference costs by utilising Gaudi’s effective processing, which will increase accessibility to AI workloads. Furthermore, dstack ensures that AI models are effectively orchestrated independent of infrastructure by supporting flexible deployment across cloud and on-premises environments.
Progressing with Scalable AI
Without the need for proprietary lock-ins, dstacks integrated approach enables you to select the optimal computing resources for each project, train models more quickly, and reduce infrastructure overhead.
Principal Benefits
- Performance Meets Accessibility: AI and data science teams can test new concepts more rapidly and maintain their lead in rapidly evolving industries to high throughput and low latency, which shorten training cycles.
- Cost-effective Scaling: You may grow AI projects without going over budget or hiring a whole operations crew to optimize your hardware by having your hardware use optimized automatically.
- Hardware Flexibility: Break free from limitations imposed by a single vendor. With Intel Gaudi and other accelerators you may already be using, dstack’s lightweight orchestration works well.
