Python SDK Reference

Auto-generated reference for the michelangelo Python package — task and workflow decorators, compute-backend plugins, trainer utilities, and workflow variable types. Start here if you're authoring pipelines with the Python SDK.

New to Michelangelo AI? See Core Concepts and Key Terms for what a task, a workflow, and Uniflow actually are before diving into the raw signatures below.

Some modules below aren't linked yet — coming soon. Two modules referenced elsewhere in this reference, the XGBoost trainer and LightningTrainer itself, aren't in the current generator config and have no page yet either; the Trainer section covers only the Torch collate-function and memory-estimation utilities so far.

Uniflow — Tasks & Workflows

The core authoring surface for pipeline tasks and workflows.

Module	Description
uniflow.core.decorator	@task and @workflow decorators
uniflow.core.task_config	Base task configuration types
uniflow.core.context	Execution context passed into tasks
uniflow.core.image_spec	ImageSpec for per-task container images
uniflow.core.io_registry	Dataset IO type registry

Uniflow — Plugins

Compute-backend plugins that extend @task with Ray, Spark, or Pandas execution.

Module	Description
uniflow.plugins.ray.task	RayTask config for Ray-backed tasks
uniflow.plugins.ray.run_config	Ray run configuration
uniflow.plugins.ray.io	Ray dataset IO
uniflow.plugins.spark.task	SparkTask config for Spark-backed tasks
uniflow.plugins.spark.io	Spark dataset IO
uniflow.plugins.pandas.io	Pandas dataset IO

CanvasFlex — Pusher

The pusher component for CanvasFlex template-driven workflows.

Module	Description
workflow.tasks.pusher.task	Pusher task definition
workflow.tasks.pusher.registry	Pusher plugin registry
workflow.tasks.pusher.exceptions	Pusher exception types

Trainer

PyTorch training utilities used with LightningTrainer (LightningTrainer itself isn't generated into this reference yet — see the note above).

Module	Description
lib.trainer.torch.data_collate_functions	Collate functions for data loading
lib.trainer.torch.utils	Trainer utilities

Native Transform

TorchScript- and ONNX-exportable transform layers for train/serve parity.

Module	Description
lib.native_transform.torch.base_layers	Base transform layer classes
lib.native_transform.torch.id_hash_tokenizer	ID hashing tokenizer layer
lib.native_transform.torch.utils	Transform utilities

Workflow Variables

Typed dataset and metadata variables for pipeline IO.

Module	Description
workflow.variables.types	DatasetVariable and related types
workflow.variables.metadata	Variable metadata types

Regenerating this reference

This section is generated from source docstrings via pydoc-markdown, configured by pydoc-markdown.yml at the repo root. If you're updating docstrings, regenerate the pages above by running bun run docs:api from the website/ directory.

pydoc-markdown isn't currently declared as a project dependency, so on a fresh checkout you may need to install it yourself (pip install pydoc-markdown) before the script succeeds.