[QAIRT] Implement QAIRT ORT->Genie workflow

[qti-kromero]

Describe your changes

Implements a complete QAIRT workflow for converting ONNX Runtime models to Genie-compatible format through three new passes:

QairtPreparation: Executes external preparation scripts to quantize and prepare HuggingFace models for QAIRT, with configurable caching and script parameters.

QairtGenAIBuilder: Converts prepared models using QAIRT GenAIBuilder API with support for:

• CPU and HTP backend targets
• Device-specific optimizations (VTCM size, HVX threads, extended UDMA)
• Model configurations (sequence lengths, multi-graph, model splits)

QairtEncapsulation: Wraps QAIRT DLC models in ONNX protobuf format with EPContext nodes, generating genai_config.json for onnxruntime-genai compatibility.

This enables end-to-end optimization of generative AI models for Qualcomm hardware accelerators.

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

[github-advanced-security]

lintrunner found more than 20 potential problems in the proposed changes. Check the Files changed tab for more details.

[qti-kromero]

dev testing complete and unit testing added - pending reviewers

[qti-kromero]

@jambayk @xiaoyu-work would it be possible to get a reviewer added to this

[jambayk]

the PR mostly looks good to me but could you use an enum with the options for backend and log_level like in

Olive/olive/passes/pytorch/rotate.py

Line 40 in 85a754a

class RotateMode(StrEnumBase):

this gives it automatic validation of the allowed values. thanks!

[Copilot]

Pull request overview

Adds a new QAIRT pipeline to Olive to support an end-to-end ORT → QAIRT (Genie) workflow, including model preparation, QAIRT GenAIBuilder compilation, and ONNX encapsulation for onnxruntime-genai compatibility.

Changes:

• Introduces three new QAIRT passes: preparation (external script runner), GenAIBuilder (CPU/HTP backends), and encapsulation (EPContext ONNX wrapper + genai_config.json).
• Adds QAIRT model handlers and new Framework/ModelFileFormat enums for QAIRT artifacts.
• Adds a new pytest suite for the QAIRT passes and updates Olive’s pass registry configuration (olive_config.json) with QAIRT entries and extra dependencies.

Reviewed changes

Copilot reviewed 12 out of 12 changed files in this pull request and generated 15 comments.

Show a summary per file

File	Description
olive/passes/qairt/preparation.py	New pass to run external preparation scripts and emit QairtPreparedModelHandler.
olive/passes/qairt/gen_ai_builder.py	New pass to compile/build QAIRT artifacts via QAIRT GenAIBuilder for CPU/HTP targets.
olive/passes/qairt/encapsulation.py	New pass to export DLC + wrap it into an ONNX EPContext model and generate genai_config.json.
olive/passes/qairt/__init__.py	Adds QAIRT passes package.
olive/model/handler/qairt.py	Adds QAIRT model handler types (QairtPreparedModelHandler, QairtModelHandler).
olive/model/handler/__init__.py	Exposes the new QAIRT model handlers from the handler package.
olive/constants.py	Adds QAIRT framework and QAIRT model file formats.
olive/olive_config.json	Registers new QAIRT passes and adds QAIRT extra dependency mapping.
test/passes/qairt/conftest.py	Adds shared fixtures for mocking QAIRT modules and model handlers.
test/passes/qairt/test_preparation.py	Unit tests for QairtPreparation behavior and subprocess streaming.
test/passes/qairt/test_gen_ai_builder.py	Unit tests for GenAIBuilder CPU/HTP behavior and validation paths.
test/passes/qairt/test_encapsulation.py	Unit tests for encapsulation, DLC discovery, ONNX generation, and genai_config creation.

[Copilot]

For HTP, set_targets([config.soc_details]) is called unconditionally, but soc_details defaults to None and validate_config doesn’t enforce it. Passing [None] into the QAIRT API is likely to fail at runtime. Add a validation/guard so HTP requires a non-empty soc_details (or skip set_targets when it’s unset and rely on QAIRT defaults).

Suggested change

	gen_ai_builder.set_targets([config.soc_details])
	if config.soc_details:
	gen_ai_builder.set_targets([config.soc_details])

[Copilot]

The log_level config option is defined but never used in this pass. Either wire it up (e.g., set the same QAIRT_LOG_LEVEL env var used in QairtGenAIBuilder) or remove it to avoid confusing users with a no-op parameter.

[Copilot]

head_size is computed as hidden_size // num_attention_heads using default values of -1 when keys are missing. This can silently produce incorrect values (e.g., -1 // -1 == 1) or raise if num_attention_heads is 0/non-int. Mirror the safer logic used in olive/passes/openvino/ov_utils.py:create_genai_config by validating both values are positive ints before dividing; otherwise set head_size to -1.

Suggested change

	genai_config["model"]["decoder"]["head_size"] = src_config.get("hidden_size", -1) // src_config.get(
	"num_attention_heads", -1
	)
	genai_config["model"]["decoder"]["hidden_size"] = src_config.get("hidden_size", -1)
	hidden_size = src_config.get("hidden_size", -1)
	num_attention_heads = src_config.get("num_attention_heads", -1)
	head_size = -1
	if isinstance(hidden_size, int) and isinstance(num_attention_heads, int) and hidden_size > 0 and num_attention_heads > 0:
	head_size = hidden_size // num_attention_heads
	genai_config["model"]["decoder"]["head_size"] = head_size
	genai_config["model"]["decoder"]["hidden_size"] = hidden_size

[Copilot]

The subprocess is invoked via the literal executable name "python", which can pick up a different interpreter than the one running Olive (e.g., in venv/conda). Use sys.executable (or equivalent) so the preparation script runs under the same Python environment and has access to the same installed dependencies.