Emit ORT-spec-compliant model packages from generate-model-package

[xiaoyu-work]

The packages produced by olive generate-model-package could not be opened by ONNX Runtime or loaded by ORT-GenAI. Verified against ORT 1.29 and a local onnxruntime-genai build, the previous output failed at the very first step: OrtModelPackageApi_CreateModelPackageContext rejected the manifest with "unknown field 'configs_dir'".

The directory layout was already fine. Four things were not:

1. manifest.json used a schema ORT rejects. ORT enforces a strict top-level whitelist, so configs_dir and producer were hard errors; components must be an object mapping name -> path, not an array; and schema_version must be a "." string, not an integer. Provenance now lives under additional_metadata.producer.

2. metadata.json is now component.json. When a manifest component entry points at a directory, ORT reads that fixed filename and nothing else.

3. genai_config_overlay.json is now a complete genai_config.json. ORT-GenAI loads Config(variant_dir, "") and explicitly refuses runtime overlays on the package path, so it has no notion of a package-level base config or an RFC 7386 merge patch. Each variant now carries a self-contained config, merged from the base, the model-level defaults, and the variant's own fields.

4. Shared config assets move from configs/ to a content-addressed shared_assets/sha256-<hex>/ directory, computed with ORT's ModelPackage_ComputeDirectoryHash algorithm. Variants reference tokenizer assets via model.tokenizer_dir = "sha256:<hex>", which is one of the few fields ORT-GenAI routes through the package resolver. Processor configs are resolved as config_path / filename instead, so they are copied into each variant directory rather than shared.

Also drops the injected model.<role>.component markers. The ORT-GenAI config parser has no component field and throws unknown_value_error on unknown keys, so those markers made every package unloadable. The comment justifying them cited a GenAI error string that does not exist in GenAI.

Fixes a latent bug along the way: a variant with no source_genai used to lose the model-level scalars that were stripped from the base. Those are now restored from a model_level_defaults fallback layer, scoped to _VARIANT_LEVEL_MODEL_KEYS so per-role filename / session_options / pipeline never leak across roles in multi-component (VLM) packages.

Verified end to end: a CPU+CUDA package generated by the CLI now opens in ORT and loads in ORT-GenAI.

Describe your changes

Checklist before requesting a review

• Add unit tests for this change.
• Make sure all tests can pass.
• Update documents if necessary.
• Lint and apply fixes to your code by running lintrunner -a
• Is this a user-facing change? If yes, give a description of this change to be included in the release notes.

(Optional) Issue link

[Copilot]

Pull request overview

This PR updates olive generate-model-package to emit ONNX Runtime model packages that conform to ORT’s strict model-package manifest/component schema and that can be loaded by both ONNX Runtime and ORT-GenAI.

Changes:

• Updates manifest.json to the ORT-accepted schema (string schema_version, component map object, provenance under additional_metadata).
• Switches per-component metadata from metadata.json to ORT-required component.json.
• Replaces genai_config_overlay.json with per-variant, fully self-contained genai_config.json, and introduces content-addressed shared_assets/sha256-<hex>/ for shared tokenizer assets.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
olive/cli/model_package.py	Reworks package writer to match ORT spec: manifest schema, component.json, per-variant full genai config, and content-addressed shared assets.
test/cli/test_model_package.py	Updates and expands tests to validate the new ORT-compliant layout and schema constraints.