Module quantization_nodes 

Low-level builders for ONNX QDQ (Quantize-Dequantize) graph primitives.

Each quantized weight becomes four graph elements:

Initializers:
  "{name}_quantized"  — INT8 tensor, same shape as original
  "{name}_scale"      — FP32 scalar
  "{name}_zp"         — INT8 scalar

Node:
  DequantizeLinear
    inputs:  ["{name}_quantized", "{name}_scale", "{name}_zp"]
    outputs: ["{name}"]          ← original name; downstream graph untouched

The DequantizeLinear op runs at inference time: output = (input - zero_point) × scale which matches the dequantize formula already used in QuantParams and QuantParamsInt4.

Structs

DequantLinearNames

Canonical names for the four graph elements that replace one FP32 initializer.

Functions

build_dequantize_linear_node

Build a DequantizeLinear NodeProto.

build_quantized_weight_tensor

INT8 tensor holding the quantized weight values.

build_scale_tensor

FP32 scale tensor.

build_zero_point_tensor

INT8 zero-point tensor.