This PR implements opset 24 versions of normalization operators with optional scale and bias inputs, allowing for optimization when scale=1.0 and bias=0.0.

Changes Made

New Opset 24 Operators

• BatchNormalization (opset 24) - scale and bias inputs now optional
• InstanceNormalization (opset 24) - scale and bias inputs now optional
• GroupNormalization (opset 24) - scale and bias inputs now optional
• LayerNormalization (opset 24) - scale and bias inputs now both optional (bias was already optional in opset 17)

Key Features

• Optional Inputs: Scale defaults to 1.0, bias defaults to 0.0 when not provided
• Smart Function Builders: Handle all 4 input combinations:
  • Both scale and bias provided
  • Only scale provided
  • Only bias provided
  • Neither provided
• Backwards Compatibility: Old versions moved to old.cc to maintain existing functionality

Implementation Details

// Example: InstanceNormalization with optional inputs
.Input(1, "scale", "The input 1-dimensional scale tensor of size C. Default value is 1.0.", "T", OpSchema::Optional, ...)
.Input(2, "B", "The input 1-dimensional bias tensor of size C. Default value is 0.0.", "T", OpSchema::Optional, ...)

Function builders include conditional logic:

if (ctx.hasInput(1) && ctx.hasInput(2)) {
  // Both scale and bias provided
} else if (ctx.hasInput(1)) {
  // Only scale provided
} else if (ctx.hasInput(2)) {
  // Only bias provided  
} else {
  // Neither provided - use defaults
}

Benefits

• Model Optimization: Enables removal of unnecessary scale/bias parameters when they equal default values
• Consistency: Makes all normalization operators behave consistently regarding optional parameters
• Backwards Compatible: Existing models continue to work unchanged

Files Modified

• onnx/defs/nn/defs.cc: Added new opset 24 definitions, removed old versions
• onnx/defs/nn/old.cc: Added old opset versions (15, 17, 21, 22) for backwards compatibility

Fixes #6552.

💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.