Overview - MKA1

Compute runs containers on the hardware available to your deployment. Use a job for work that finishes, such as training, and a service for a persistent process, such as a model server. Compute schedules and observes the workload. Your container supplies the training or serving software.

Choose a workflow

Goal	Guide
Store weights and training data	Repositories
Generate training data	Datasets
Run a finite training workload	Run a fine-tune job
Train against a reward function	Reinforcement learning
Keep a model endpoint running	Deploy a model server
Inspect capacity, workloads, secrets, and volumes	Manage resources

How a workload runs

Select hardware from your deployment’s catalog and request a quote.
Supply a container image, command, and configuration.
Put credentials in Compute secrets and reference them from the workload.
Create a job or service and inspect its status, logs, and events.
Publish outputs to a repository or supported persistent storage.
Terminate persistent services when you finish using them.

A quote reports price and availability; it does not reserve hardware. Provider capabilities determine which storage and scaling options are available. Your organization must have Compute access enabled.

Keep artifacts separate from the runtime

The container’s ephemeral disk is not an artifact repository. Training workloads should publish checkpoints and final outputs explicitly. Repositories gives jobs and model servers a shared identifier for those artifacts.

Follow the feature status

Individual guides identify early-access or upcoming functionality. Those labels describe the specific workflow; they do not imply that every provider supports every option.

⌘I