Preparing for Distribution

This documentation describes how to create a distributable version of the model backend for sharing inference code.

Before Distribution

Before distributing a Packflow project, it is important to ensure that the Inference Backend has been properly validated and tested. This ensures that the backend functions as expected and meets the necessary requirements for deployment.

It is also strongly recommended to complete the following files in the project structure to provide necessary context and information for downstream users:

  • MODEL_CARD.md: Document model details, limitations, and ownership.

  • README.md: Provide usage instructions and relevant information for developers and contributors.

  • LICENSE.txt: Specify the applicable licensing information for the project.

Additional Requirements:

  • Complete packflow.yaml: Ensure that the Packflow configuration file is fully populated with accurate metadata and configuration details.

  • Specify dependencies in requirements.txt: List all necessary Python dependencies to ensure that the environment can be properly set up for the Inference Backend.

These files are created as part of the Packflow project structure when initializing a new project with the packflow create command.

Packflow Config File

After initializing a project with the packflow create command, a file named packflow.yaml will be automatically generated in the project structure. This file contains information that Packflow can read to load the project, and presents a natural integration surface for other systems to load and interact with Packflow-built analytics.

Contents

The purpose of packflow.yaml is to track:

  1. Project metadata, including name, version, description, and maintainers.

    • This information could be used as a part of a DevOps pipeline or workflow designed to efficiently import Packflow projects into a system.

Note

For instance, a DevOps pipeline could send automated emails to warn the maintainers of an issue with running the project, or to alert them of a vulnerability detected in one of the project’s dependencies.

  1. Loader configuration, including python_version, inference_backend, and loader.

    • These configuration fields specify how an Inference Backend may be loaded from the Packflow project.

  2. Custom configuration/metadata fields

    • If there are any other project-scope configurations or metadata that needs to be tracked, arbitrary configuration values are accepted here.

Note

Please also consider Backend Configuration for configurations specific to Packflow’s preprocessors or Inference Backend, and for defining custom configurations and behaviors in the InferenceBackend controlled by the backend config.

Creating a Package

After ensuring that the Packflow project is properly validated, tested, and documented, create a distributable package of the project using the packflow export [PROJECT_PATH] command.

packflow export [PROJECT_PATH]

The [PROJECT_PATH] argument is optional; if not provided, the current working directory will be used as the project path.

This will create a zipped archive of the Packflow project in the current working directory, which can then be shared and deployed across different environments.

Note

Creating a new package should equate to a new ‘release’ of this project, and it is recommended to create packages based on Git branches or tags for versioning purposes.