Express deployments for model serving endpoints

Express deployment reduces deployment time for custom model serving endpoints. For eligible models on CPU or GPU serving compute, Azure Databricks automatically prepares and reuses compatible model environments when you create or update an endpoint. No additional registration parameter or deployment-mode selection is required.

Requirements

Express deployment supports custom models on CPU and GPU serving endpoints.

The model must be logged with MLflow from Azure Databricks compute, either classic compute running Databricks Runtime or serverless compute.

Azure Databricks also checks whether the model's recorded runtime, Python version, and dependencies are compatible with the selected serving compute. See Configurations that can affect deployment time.

Deploy a model

Log and register your model, then create or update a serving endpoint with CPU or GPU compute. You don't need to prepare the environment separately for eligible models.

Include the dependencies your model needs when you log it with MLflow. Express deployment uses the model's recorded runtime and dependencies, rather than a copy of your running notebook session. See Deployment container and dependencies.

What to expect during deployment

The first deployment of a model can take additional time while Azure Databricks prepares its environment. Later deployments can reuse a compatible prepared environment. Deployment time also depends on model size, model loading, and compute availability.

Configurations that can affect deployment time

When Express Deployment cannot be applied, Azure Databricks uses a deployment process that can take longer. Configurations that can prevent express deployment include:

Configuration Cases that can prevent express deployment
Runtime and Python Missing runtime or Python-version metadata, or a runtime or Python version that isn't compatible with the selected CPU or GPU serving compute.
Model environment A Conda-only environment or a dependency on PySpark.
Dependency requirements Unsupported entries in the model's pip requirements, such as editable installs, Git or URL dependencies, package-index overrides, or bundled wheels without a content hash.
GPU dependencies Unsupported GPU-specific requirements, such as CUDA-tagged package versions. For models logged on serverless compute and served on GPU, declared torch, torchvision, and torchaudio requirements must be exact version pins matching the GPU serving environment.
Model size Missing model-size metadata or model artifacts that exceed express deployment size limits.
Environment-file size Python environment or requirements files that exceed the supported file-size limit.
Workspace configuration A customer-managed key for managed services.

You don't need to select a different deployment mode.

GPU example with explicit environment preparation

Explicit environment preparation with env_pack remains supported for workflows that need to package the notebook environment during model registration. The following notebook uses this workflow to deploy a GPU reranker. Eligible CPU and GPU models can use express deployment without this parameter.

Run the notebook on AI Runtime with an A10 GPU and the AI environment, version 5. The notebook installs the required MLflow and Databricks SDK versions and registers the model with env_pack=EnvPackConfig(name="databricks_model_serving").

EnvPackConfig installs the model's dependencies before packaging by default. For workspaces without internet access or with custom libraries, you can set install_dependencies=False when the required dependencies are already available in the notebook environment.

Express reranker starter notebook

Get notebook