Module fusion 

Kernel fusion: merges adjacent classified patterns into fused operations.

After pattern classification, fuse_patterns() scans adjacent patterns and merges compatible sequences (e.g., Conv+BatchNorm, Add+ReLU).

Enums

FusedActivation

Fused activation function appended to a base operation.

FusedPattern

A pattern that may be fused from one or more classified patterns.

Functions

fuse_patterns

Greedy adjacent fusion of classified kernel patterns.

input_tensor_names

Returns the input tensor names of a pattern.

output_tensor_names

Returns the output tensor names of a pattern.

tensors_connect

Returns true if any output tensor name of producer matches any input tensor name of consumer, indicating data flows between them.