Integrate CGIR workflow (export, perf, eval) into winml-cli

[Zhenchao Ni (zhenchaoni)]

Summary

Integrates CGIR export, configuration, build, compilation, performance measurement, and output comparison into WinML CLI.

Three inference workflows are supported:

Workflow	Input	Runtime / Backend
Offline CGIR	Standalone MLIR	Windows ML Runtime / CGC
WinMLCG EP	ONNX	ONNX Runtime / WinMLCG EP
Online CGC compilation	ONNX	Windows ML Runtime / CGC

Examples below use ResNet50 and the installed preview Runtime/EP packages. --no-sync preserves those packages in the local environment.

Export ONNX and CGIR

# Export Hugging Face model to ONNX
uv run --no-sync winml export -m microsoft/resnet-50 -o temp\demo-cgc\resnet50.onnx --overwrite

# Export Hugging Face model directly to standalone CGIR
uv run --no-sync winml export -m microsoft/resnet-50 -o temp\demo-cgc\resnet50.mlir --target cgir --overwrite

# Alternatively, convert an existing ONNX model to CGIR
uv run --no-sync winml export -m temp\demo-cgc\resnet50.onnx -o temp\demo-cgc\resnet50.mlir --target cgir --overwrite

Performance: Three Inference Workflows

# 1. Standalone MLIR through Windows ML Runtime
uv run --no-sync winml perf -m temp\demo-cgc\resnet50.mlir --runtime winml-runtime --device gpu --output temp\demo-cgc\perf-mlir.json

# 2. ONNX through ONNX Runtime + WinMLCG EP
uv run --no-sync winml perf -m temp\demo-cgc\resnet50.onnx --runtime winml-ort --ep winmlcg --device gpu --output temp\demo-cgc\perf-winmlcg.json

# 3. ONNX through Windows ML Runtime with online CGC compilation
uv run --no-sync winml perf -m temp\demo-cgc\resnet50.onnx --runtime winml-runtime --backend cgc --device gpu --iterations 100 --output temp\demo-cgc\perf-runtime-cgc.json

perf defaults to automatic runtime selection: MLIR selects winml-runtime; ONNX selects winml-ort. Explicit runtime selection remains available. With winml-runtime, ONNX defaults to the CGC backend.

Compare Against the Original ONNX

Use ORT CPU as the reference and compare the outputs from each workflow. These examples use one sample for a quick check; increase --samples for broader coverage.

# MLIR vs original ONNX
uv run --no-sync winml eval --mode compare -m temp\demo-cgc\resnet50.mlir --runtime winml-runtime --device gpu --reference temp\demo-cgc\resnet50.onnx --reference-device cpu --samples 1 --output temp\demo-cgc\compare-mlir.json

# WinMLCG EP vs original ONNX
uv run --no-sync winml eval --mode compare -m temp\demo-cgc\resnet50.onnx --runtime winml-ort --ep winmlcg --device gpu --reference temp\demo-cgc\resnet50.onnx --reference-device cpu --samples 1 --output temp\demo-cgc\compare-winmlcg.json

# Runtime online CGC vs original ONNX
uv run --no-sync winml eval --mode compare -m temp\demo-cgc\resnet50.onnx --runtime winml-runtime --backend cgc --device gpu --reference temp\demo-cgc\resnet50.onnx --reference-device cpu --samples 1 --output temp\demo-cgc\compare-runtime-cgc.json

Compile a WinMLCG EPContext Model

uv run --no-sync winml compile -m temp\demo-cgc\resnet50.onnx --ep winmlcg --device gpu -o temp\demo-cgc\resnet50-winmlcg-ctx.onnx --overwrite

Generate Configs and Build

--backend cgc prepares the model and converts it to CGIR. --ep winmlcg prepares ONNX for the WinMLCG EP without adding CGIR conversion. Use one selector, not both.

# Generate and review a CGC build configuration
uv run --no-sync winml config -m temp\demo-cgc\resnet50.onnx --backend cgc --device gpu -o temp\demo-cgc\config-cgc.json --overwrite

# Build using that configuration
uv run --no-sync winml build -c temp\demo-cgc\config-cgc.json -m temp\demo-cgc\resnet50.onnx -o temp\demo-cgc\build-cgc

# Or auto-generate the configuration and build in one command
uv run --no-sync winml build -m temp\demo-cgc\resnet50.onnx --backend cgc --device gpu -o temp\demo-cgc\build-cgc-auto

# Generate a configuration for WinMLCG EP
uv run --no-sync winml config -m temp\demo-cgc\resnet50.onnx --ep winmlcg --device gpu -o temp\demo-cgc\config-winmlcg.json --overwrite

# Prepare ONNX for WinMLCG EP with an auto-generated configuration
uv run --no-sync winml build -m temp\demo-cgc\resnet50.onnx --ep winmlcg --device gpu -o temp\demo-cgc\build-winmlcg

These configurations apply CGC compatibility rules and default FP16 preparation. Use --no-quant to preserve the existing ONNX precision, and --rebuild to regenerate existing build artifacts. Review output accuracy, particularly when lossy compatibility rules such as cubic-to-linear Resize are applied.

Standalone Compatibility Optimization

Compatibility rules can also be selected individually before export or inference:

uv run --no-sync winml optimize -m temp\demo-cgc\resnet50.onnx -o temp\demo-cgc\resnet50-compatible.onnx --disable-ort-graph-optimization --enable-omit-empty-resize-inputs --enable-tile-repeats-initializer-to-constant --enable-fold-scalar-initializer-casts --enable-resize-tf-half-pixel-for-nn-to-asymmetric

Select a Specific GPU

# List devices and their LUIDs
uv run --no-sync winml sys

# Replace the placeholder with a LUID from winml sys
uv run --no-sync winml perf -m temp\demo-cgc\resnet50.mlir --runtime winml-runtime --device gpu --device-luid "<LUID>"

Runtime CGC fails explicitly if the selected adapter cannot be resolved; it does not silently switch to another GPU.

[Qiong Wu (qiowu) (DingmaomaoBJTU)]

CI follow-ups for the reviewed commit. These comments identify failing checks and test updates; they do not assert a demonstrated production inference regression.

[Qiong Wu (qiowu) (DingmaomaoBJTU)]

Reviewed this integration in the context of the earlier mirror PR review, including the retained adapter/scalar/error-handling fixes, added CGC constant/DQ/GridSample/Identity rewrites, exporter publication rollback, and Runtime/build/eval routing. No blocking findings identified. All reported PR CI checks pass and git diff --check passed. Seven Foundry-binding tests passed locally in an isolated harness (two missing pytest-timeout configuration warnings). The broader local exporter/runtime test attempts could not collect under that minimal harness because package exports/native dependencies were not available; those attempts are not counted as passing. Native Windows ML/Foundry inference, rewrite numerical suites and the full unit suite were not rerun locally.