Module graph_rewrite 

Graph Rewrite Passes for INT8 Quantization (ADR-091 Phase 3)

This module implements four critical graph optimization passes:

• GR-1: fuse_batchnorm_to_conv - Absorb BatchNorm into Conv weights/bias
• GR-2: fuse_zp_to_bias - Pre-compute zero-point correction in bias
• GR-3: insert_qdq_nodes - Insert Quantize/Dequantize nodes at boundaries
• GR-4: fuse_relu/fuse_hardswish - Merge activations into preceding ops

Structs

ComputationGraph

Computation graph for optimization passes

GraphNode

Graph node representing a single operation

Enums

NodeParams

Parameters for different node types

NodeType

Computation graph node types

Functions

fuse_batchnorm_to_conv

GR-1: Fuse BatchNorm parameters into Conv weights and bias

fuse_hardswish

GR-4: Fuse HardSwish activation using LUT

fuse_relu

GR-4: Fuse ReLU activation into preceding convolution

fuse_zp_to_bias

GR-2: Fuse zero-point correction into bias

generate_hardswish_lut

Generate HardSwish LUT for INT8 quantized values

insert_qdq_nodes

GR-3: Insert Quantize/Dequantize nodes at INT8 subgraph boundaries