Skip to main content

​
Introduction

Models are the core AI artifacts in Edge Inference Services. Each model encapsulates a trained machine learning algorithm with its weights, metadata, and runtime requirements. Models live inside organizations and can be imported, versioned, benchmarked, and deployed to nodes. Models were previously called Model Hub. You manage models at Inference > Deploy > Models in the ZEDEDA console. The Models list shows each model with its name, description, ML framework, organization, tags, and last-updated date.

​
Prerequisites

  • To import, update, or delete models, you have the Organization Admin, Data Scientist, or Super Admin role. The Organization Observer, MLOps Engineer, MLOps Admin, and Super Observer roles can view models. See Organizations, Users, and Roles.
  • Access to at least one organization.
  • For Jupyter Notebook upload: Python 3.7 or later and the ZEDEDA Edge AI Client installed.
  • For deployment: A node running ZEDEDA Edge Kubernetes Service and network connectivity between Edge Inference Services, Edge Infrastructure Services, and the node.

​
View Models

  1. Go to Inference > Deploy > Models.
  2. The Models list displays all models in your accessible organizations.
From this view you can:
  • Filter by framework, organization, or tags using the filter controls.
  • Search models by name using the search bar.
  • Select a model card to open its detail page.

​
View Model Details

  1. Go to Inference > Deploy > Models.
  2. Find and select a model to open its detail page.
The detail page displays the following information. Metadata:
  • Model name, version, and description.
  • Creation and last-updated timestamps.
  • Creator and owner.
  • Model stage. See Model Stages.
  • Tags for categorization.
Metrics:
  • Training accuracy, loss, and custom metrics.
  • Validation performance.
  • Comparison with previous versions.
Technical specifications:
  • ML framework and version (for example, TensorFlow, PyTorch, XGBoost).
  • Model input shapes and data types.
  • Model size and architecture details.
Artifacts:
  • Model files (for example, .pb, .pth, .onnx).
  • Documentation files and model cards.
  • Test datasets.
  • Preprocessing and postprocessing scripts.

​
Import a Model

You can import models into Edge Inference Services using three methods.

​
Upload from the console

Use this method for direct file upload. A Local (file upload) provider must be configured and active in the target organization. The upload runs as an asynchronous import job, and the model appears in the model list when the job completes.
  1. Go to Inference > Deploy > Models.
  2. Select Upload from the toolbar.
  3. Select an organization to place the model in.
  4. Provide the model name, description, and framework information.
  5. Upload the model files.
  6. Select Save.

​
Import from an external provider

Use this method to pull models from a configured external provider: Hugging Face, Azure ML, MLflow, Amazon S3, Azure Blob Storage, Amazon SageMaker, NVIDIA NGC, or Qualcomm AI Hub.
  1. Go to Inference > Deploy > Providers and confirm that the provider is configured. See Providers.
  2. Go to Inference > Deploy > Models.
  3. Select Import from the toolbar.
  4. Select the provider and enter the source model identifier.
  5. Select the target organization. The provider must belong to the target organization; otherwise the import is rejected.
  6. Select Import. The import runs asynchronously and appears in the model list when complete.

​
Upload from a Jupyter Notebook

Use this method to log models directly from a training run using the ZEDEDA Edge AI SDK.
  1. Install the ZEDEDA Edge AI Client in your Python environment.
  2. Use the SDK’s model logging API during or after training to upload the model, capture metadata, and associate it with an organization.
Refer to the ZEDEDA Edge AI SDK documentation for the required API calls and configuration.

​
Model Stages

The stage indicates where a model version sits in its lifecycle. Each model version has its own stage. Newly imported versions start in the Development stage. You can promote or demote a version by editing that version.

​
Update a Model

You can update a model’s name, description, or tags after it is imported. You can update a model version’s description, tags, or stage.
  1. Go to Inference > Deploy > Models and open the model detail page.
  2. Select the Edit icon.
  3. Update the fields.
  4. Select Save.
The organization and framework cannot be changed after a model is created.

​
Delete a Model or Model Version

Before deleting a model, verify that it is not referenced in an active deployment or benchmark job. To delete a model version:
  1. Open the model detail page.
  2. Navigate to the version history section.
  3. Select the version to delete.
  4. Select Delete Version and confirm.
To delete an entire model:
  1. Open the model detail page.
  2. Select Delete Model from the actions menu.
  3. Confirm the deletion. This action cannot be undone and permanently removes all associated data.

​
Benchmark a Model

Before deploying a model to production, validate its performance on real edge hardware using the Evaluation feature.
  1. Open the model detail page.
  2. Select Benchmark from the actions menu.
  3. Select one or more target device pools. Each run uses one available cluster from each selected device pool.
  4. Start the benchmark run.
See Inference Model Benchmarking for the full benchmarking workflow.

​
Deploy a Model

After benchmarking, you can deploy the model to nodes or Kubernetes clusters.
  1. Open the model detail page.
  2. Select Deploy from the actions menu.
  3. The Model Deployment Wizard opens. Follow the wizard steps to select the target node or cluster, configure the application layer (API exposure, data transformations), and submit the deployment.
Deploying to a Kubernetes cluster creates a deployment request that routes through a GitOps-based approval pipeline. See Submit Model Deployment Requests to track and manage that request. See Approve and Validate Models for reviewing and approving active deployments.

​
In This Section