This PR adds a newIODataTypeConverter surgeon to theGraphSurgeries pass, providing the same functionality as the standaloneOnnxIODataTypeConverter pass but integrated into the graph surgeries framework.

Changes Made

New Surgeon Implementation

• AddedIODataTypeConverter class inheriting fromProtoSurgeon
• Implements full datatype conversion functionality:
  • Input/output datatype conversion with Cast node insertion
  • Name pattern matching using regex
  • Comprehensive datatype validation
  • Proper I/O mapping for graph transformation

Key Features

• Parameter Compatibility: Uses the same parameters asOnnxIODataTypeConverter:
  • source_dtype: Source datatype integer value
  • target_dtype: Target datatype integer value
  • name_pattern: Regex pattern for selective conversion (default: "logits")
• Graph Transformation: Adds Cast nodes at graph boundaries to convert datatypes
• Validation: Validates ONNX datatype values and rejects invalid types
• Integration: Seamlessly works within the existing GraphSurgeries framework

Usage Example

{"type":"GraphSurgeries","surgeries": [        {"surgeon":"IODataTypeConverter","source_dtype":10,"target_dtype":1,"name_pattern":"logits"        }    ]}

This converts all inputs/outputs matching "logits" from FLOAT16 (10) to FLOAT (1).

Documentation and Testing

• Added comprehensive documentation with transformation examples
• Included unit tests covering both valid and invalid datatype scenarios
• Verified equivalence with the originalOnnxIODataTypeConverter pass

This implementation provides a migration path for users to move from the standalone pass to the more flexible GraphSurgeries framework, supporting the long-term goal of consolidating ONNX transformations.

Fixes#1928.

💬 Share your feedback on Copilot coding agent for the chance to win a $200 gift card! Clickhere to start the survey.