Add GenAI runtime profile model variants

[Tianlei Wu (tianleiwu)]

Describe your changes

Add an Olive post-processing pass for ONNX Runtime GenAI memory profiles. A profile can update only genai_config.json, or derive an alternate KV-cache decoder graph while hardlinking the rest of the generated package and preserving the base external-data file.

This replaces #2663 and builds on the composite ModelBuilder support now present on main. It depends on microsoft/onnxruntime-genai#2622 for runtime profile fields and the KV-cache graph transformer.

Key changes

• Add and register GenAIModelRuntimeProfiles for configuration-only and alternate-decoder profiles.
• Extend ModelBuilder structured configuration forwarding and CPU-embedding split handling needed by the combined Qwen package.
• Generate deterministic model_<profile-id>.onnx filenames and strip build-only kv_cache settings from emitted runtime configuration.
• Validate profile IDs, eligibility, filenames, and required graph-generation settings.
• Add focused ModelBuilder and runtime-profile pass tests.

Validation

• Direct focused runtime-profile pass test passed for both graph-generating and config-only modes.
• Real Qwen3.8-27B package validation created model_32gib.onnx, retained the same hardlinked model.onnx.data, and emitted the requested runtime overlay.
• python -m ruff check ... and python -m ruff format --check ... passed on changed Python files.
• scripts/format_json.py --linter-mode ... olive/olive_config.json passed.
• Full pytest collection was not available locally because the environment lacks the optional peft test dependency; the focused test body was invoked directly.

Checklist before requesting a review

• Add unit tests for this change.
• Make sure focused tests pass.
• Update documents if necessary. (Runtime/profile documentation is in Support memory-based runtime model profiles onnxruntime-genai#2622.)
• Lint and apply fixes to the changed files.
• Is this a user-facing change? The new GenAIModelRuntimeProfiles pass enables memory-tiered GenAI packages.

(Optional) Issue link

• Supersedes Support current GenAI composite model exports #2663
• Depends on Support memory-based runtime model profiles onnxruntime-genai#2622

[Copilot]

Copilot review overview

🟡 Changes recommended

Unresolved correctness, validation, path-safety, and hardlink-isolation issues remain.

Get a fresh assessment by requesting another Copilot review.

Review effort: Lite
Findings: 1 · 4

Open (5)

• Reject absolute and traversing source filenames · New
• Validate eligibility ranges and overlaps · New
• Validate overlay is a mapping · New
• Break hardlinks before generating model variants · New
• Break config hardlink before writing · New

What changed in this PR

Adds GenAI runtime profiles for configuration overlays and alternate KV-cache decoder graphs, with supporting ModelBuilder updates.

Changes:

• Adds and registers GenAIModelRuntimeProfiles.
• Extends structured builder options and CPU embedding splitting.
• Adds focused tests and configuration registration.

File	Summary
test/​passes/​onnx/​test_model_builder.py	Tests structured configuration and embedding splitting.
test/​passes/​onnx/​test_genai_runtime_profiles.py	Tests runtime profile and decoder variant generation.
olive/​passes/​onnx/​model_builder.py	Adds structured options and CPU embedding support.
olive/​passes/​onnx/​genai_runtime_profiles.py	Implements profile validation, overlays, package copying, and decoder variants.
olive/​olive_config.json	Registers the new pass.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.