Repository Structure and CI Workflows#

This page explains how the source code is organized and how the continuous-integration workflows are structured for each component of yet-another-onnx-builder (yobx).

Overview#

The repository is organized as a set of mostly-independent optional sub-packages under yobx/. Each sub-package has its own unit-test directory under unittests/ and typically a dedicated CI workflow. The table below summarizes the mapping:

Library

Source

Tests

CI Workflows

Core / Builder

yobx/xbuilder/, yobx/xoptim/, yobx/xshape/, yobx/container/, yobx/helpers/, yobx/reference/, yobx/translate/

unittests/ (excluding library sub-directories)

core_tests.yml

scikit-learn (and xgboost, lightgbm)

yobx/sklearn/

unittests/sklearn/

sklearn_tests.yml

PyTorch / Transformers

yobx/torch/

unittests/torch/

torch_tests.yml, ci_transformers_dev.yml, ci_transformers_releases.yml

TensorFlow

yobx/tensorflow/

unittests/tensorflow/

tensorflow_tests.yml

On top of the per-library workflows there are cross-cutting quality checks that run on every push:

Workflow

Purpose

build.yml

Builds and uploads the Python package (python -m build).

docs.yml

Builds the Sphinx documentation (triggered by .rst / .md changes only). TensorFlow gallery examples are built first in a dedicated single-threaded pass (sphinx_gallery_conf.filename_pattern=/tensorflow/), then the full documentation is built in a second pass.

mypy.yml

Static type-checking of the whole yobx/ package.

style.yml

Linting with ruff.

spelling.yml

Spell-checking of documentation strings with codespell.

pyrefly.yml

Experimental type-analysis with pyrefly.

Core / Builder#

The core of the library lives in several tightly-coupled sub-packages:

  • yobx/xbuilder/ — GraphBuilder, the main API for constructing and optimizing ONNX graphs. See GraphBuilder for a full description.

  • yobx/xoptim/ — pattern-based graph optimizer and all optimization patterns (patterns/, patterns_ort/, patterns_ml/, patterns_exp/). See Available Patterns.

  • yobx/xshape/ — symbolic shape-expression system used internally by GraphBuilder. See ShapeBuilder.

  • yobx/builder/ — lightweight builder helpers and the OnnxScriptGraphBuilder bridge.

  • yobx/container/ — ModelContainer for grouping several ONNX models that share initializers.

  • yobx/helpers/ — general ONNX and runtime utilities shared by all sub-packages.

  • yobx/reference/ — reference evaluators (onnxruntime, torch).

  • yobx/translate/ — translates ONNX graphs back to executable Python.

Tests#

Core tests live directly under unittests/. Each library has its own subfolder and CI file.

On this page

This Page