This pull request introduces support for a new "Model Package" feature, which enables handling and packaging multiple model variants (e.g., for different hardware targets or deployment configurations) as a single logical unit. It adds a new ModelPackageModelHandler, a corresponding CLI command, and extends the pass system to recognize and process model packages. Several parts of the codebase are updated to support this new functionality, including improvements to file handling and documentation.

Model Package Support

• Added ModelPackageModelHandler in olive/model/handler/model_package.py, allowing the management of multiple deployment variants of the same model as a single package. This handler enforces that all variants are of the same type and provides utilities for serialization and variant iteration.
• Registered ModelPackageModelHandler in the model handler registry and __init__.py to make it available throughout Olive. [1] [2]
• Added a new ModelPackage pass to the configuration and documentation, supporting the creation and management of model packages. [1] [2]
• Added a new CLI command (generate-model-package) for model package operations and updated documentation to describe its usage. [1] [2] [3]

Pass System Enhancements

• Updated the Pass base class to add _accepts_model_package_model and _skip_additional_files_carry_forward flags, enabling passes to control how model packages are processed and how files are handled.
• Modified the pass execution logic to automatically run passes independently on each model package variant unless the pass opts in to handle the package as a whole. The output is then reassembled as a new ModelPackageModelHandler. [1] [2]
• Improved file carry-forward logic to support directories and to allow passes to skip this step if needed. [1] [2]
• Updated the EPContextBinaryGenerator pass to optionally generate context binaries for multiple provider options and return a model package when a list of options is provided. [1] [2] [3]

Model and File Handling Improvements

• Enhanced save_model in olive/cache.py to properly save model packages, copying only relevant subdirectories and updating paths and manifest information accordingly.
• Updated output directory logic in engine.py to better distinguish between file and directory outputs, ensuring correct artifact placement for model packages. [1] [2]

Miscellaneous

• Adjusted node exclusion pattern for matmul quantization in optimize.py.

These changes collectively enable Olive to support advanced deployment workflows where a single model package can encapsulate multiple optimized variants, streamlining model management and deployment for heterogeneous environments.

ORT reference: microsoft/onnxruntime#27786