Expand description
Graph-level operations for quantized ONNX models.
Three responsibilities:
- QDQ transform — replace FP32 initializers with INT8 + DequantizeLinear
- Connectivity validation — walk the graph and verify every edge resolves
- Opset management — ensure the model declares opset ≥ 13
Structs§
- Connectivity
Report - Result of a graph-connectivity check.
- QdqWeight
Input - One weight to convert: FP32 initializer → INT8 + DequantizeLinear block.
Functions§
- apply_
qdq_ transform - Replace FP32 weight initializers with INT8 quantized equivalents + DequantizeLinear nodes.
- ensure_
opset_ version - Ensure the default ONNX domain opset is at least
min_version. - validate_
graph_ connectivity - Walk the graph and verify every node input resolves to something.