Skip to main content
Version: Cloud

Create a Deployment

A Deployment represents an instance of a machine learning model. Every Deployment has an owner, who assumes responsibility for its management. Initially, the user creating the deployment automatically becomes its owner. Creating Deployments is a straightforward process that can be completed in a few simple steps, using the Deployment API, Python Client, or UI. Some steps are unique to specific Deployment types.

tip

For large Repositories, we suggest creating Deployments using the Python Client to speed up the process.

Prerequisites

  • The model (and, if applicable, the explainer and transformer) is available in a Git repository or another accessible location. When using a Git repository, ensure it meets all repository requirements.
  • When required, credentials have been added to the Workspace.

Repository

Though it’s optional, we recommend linking a Git repository. Either link a repository, connect an existing linked Repository, or choose not to link one.

If you link a repository, select the branch and commit to associate with the Deployment version. If your model, explainer, transformer folders, or reference configuration aren’t in the root directory of the selected commit, uncheck Use root folder and select the folder that contains them.

Deployment

Configure the details of your Deployment. Select the associated use case. If you don’t select a use case, one is created for you. Select any guardrails that should apply to the Deployment.

If you’ve connected a Repository that contains metadata, you can retrieve and review it. If you haven’t connected a Repository, you can provide metadata in JSON format.

If you’re creating a managed Deployment, you can change the deployment service. Clear Use default Deployment service and select your preferred service. Workspace owners can change the default Deployment service on the Workspace Integrations page.

Model (managed Deployments)

Select the model framework you used to train the model. See supported framework versions for supported frameworks, versions, and examples. Alternatively, you can deploy a custom Docker image.

If you haven’t connected a Repository, provide a reference to the model location in JSON format.

For details on advanced model options, see advanced Deployment options.

Explainer (managed Deployments)

A standard explainer is available for select model types. Alternatively, you can deploy a trained explainer or choose no explainer. See supported framework versions for supported frameworks, versions, and examples.

If you want to use a trained explainer and haven't connected a Repository, provide a reference to the explainer location in JSON format.

For details on advanced explainer options, see Advanced Deployment options.

Transformer (managed Deployments)

If you have trained a transformer, select the Custom Docker option to deploy a transformer. See supported framework versions for supported frameworks, versions, and examples.

If you haven't connected a Repository, provide a reference to the transformer location in JSON format.

For details on the advanced transformer options, see Advanced Deployment options.

Connection (external Deployments)

Provide the URL where your external model is accessible. Select your preferred authentication method and enter your credentials. You can optionally run a connection check to confirm Deeploy can access the model before continuing. In addition, this check is always ran automatically during deployment.

For more information on the authentication, see External Deployment authentication.

Compliance

Add the compliance templates that are relevant for your Deployment. You can complete them right away, or any time after deployment.

Summary

Check the configuration of your Deployment before continuing.

Request approval for your changes from other Workspace members. Click Create pending version, and a pending version of the Deployment will be available on the Versions page awaiting approval. The Deployment owner can deploy the pending version once approved.