Skip to main content

Module fusion

Module fusion 

Source
Expand description

Operator fusion pass.

Fuses adjacent operations in the compute graph into single fused operations to reduce memory traffic and kernel launch overhead.

Supported fusion patterns:

  • Conv2D + Bias + Activation โ†’ FusedConv2d with activation attribute
  • MatMul + Bias (Add) โ†’ Gemm (with beta=1)
  • ElementWise + Activation โ†’ FusedElementWise with activation

Structsยง

OperatorFusion
Fuses adjacent operations in the compute graph into single fused ops.