quark-onnx-result-validator

You can verify Quark ONNX quantization output by checking auxiliary file copying, initializer byte identity, model metadata consistency, and quantization signal presence. This Skill automates post-quantization inspection of model.onnx and model.onnx_data files.

Initializer byte identity validation works by reading external-data bytes and comparing weights that should remain unchanged after quantization. It ensures excluded initializers stay byte-identical between the source and quantized output directories.

QDQ pattern inspection is a quantization signal detection process that produces a header-only summary. It identifies QDQ patterns, custom Quark ops, node types, and dtype distributions within the quantized ONNX model graph.

Yes, the metadata comparison ignores Quark-injected opset domains when checking model metadata and graph I/O signatures. This ensures validation focuses on original model properties rather than framework-specific quantization artifacts.

Yes, you can validate external ONNX data files alongside the main model. The Skill performs optional external-data byte reads to compare model.onnx and model.onnx_data files across source and output directories.

Auxiliary files may be missing if non-model support files were not copied from the source directory to the quantized output directory. This Skill checks auxiliary file alignment to confirm all required support files are present.