feat: add --allow-unsupported-nodes to perf/build/eval/run

[xieofxie]

Problem

winml perf -m depth-anything/Depth-Anything-V2-Small-hf --task depth-estimation --ep openvino --device gpu aborts at the optimize/analyze loop:

RuntimeError: Unsupported nodes persist after N analyze pass(es): ['OP/ai.onnx/Resize']

The static analyzer flags nodes the target EP doesn't fully support (e.g. Resize on OpenVINO GPU), so the build raises — even though the EP can still run the model (CPU fallback for those nodes).

Change

Add an opt-in --allow-unsupported-nodes flag that turns the hard failure into a warning and lets the build proceed. The gate lives at the single raise in build/common.py:

if analysis is not None and analysis.has_errors:
    message = f"Unsupported nodes persist after {analyze_iterations} analyze pass(es): ..."
    if allow_unsupported_nodes:
        logger.warning("%s. Continuing anyway ...", message)
    else:
        raise RuntimeError(message)

The flag is a shared CLI option (cli_utils.allow_unsupported_nodes_option()) applied to every command that builds a model via WinMLAutoModel.from_*:

Command	Path	Status
winml perf	BenchmarkConfig → from_pretrained/from_onnx; --module → build_hf_model	✅ verified e2e
winml build	_run_optimize_stage → run_optimize_analyze_loop; module mode → build_hf_model	✅ verified e2e
winml eval	WinMLEvaluationConfig.allow_unsupported_nodes → from_pretrained	✅
winml run	InferenceEngine.load → _load_from_hf → from_pretrained	plumbed (command currently disabled)

ONNX / skip_build load paths don't run the analyzer, so they're untouched. Default is False everywhere — strictly opt-in. Named --allow-unsupported-nodes (not a generic --force) since it gates exactly this analyzer check.

Verification

• winml perf ... --ep openvino --device gpu --allow-unsupported-nodes → builds, compiles on OV GPU, benchmarks (~15.9 ms avg).
• winml build ... --ep openvino --device gpu --allow-unsupported-nodes → logs the warning (['OP/ai.onnx/Resize']) and completes.
• New/updated unit tests cover the gate (build/test_common.py), and flag passthrough for build/eval/run; full commands + inference + build + eval suites pass (871 tests).

🤖 Generated with Claude Code

[xieofxie]

Note: although AI says (CPU fallback for those nodes) , the op profiler find that gpu supports it as interpolate type