Support Optional shapes

[yuslepukhin]

This pull request introduces a new mechanism for handling tensor shape information in ONNX Runtime, improving the representation and querying of shape and symbolic dimension data. The changes add support for tensors that may not have shape information, update APIs to reflect this, and refactor internal shape handling logic for improved consistency and extensibility.

API and Core Logic Improvements

• Added a new API method TensorTypeAndShape_HasShape to OrtApi and corresponding C++ wrapper HasShape() in TensorTypeAndShapeInfoImpl, allowing users to check if a tensor has shape information. [1] [2] [3]
• Updated all shape-related API implementations (e.g., GetDimensionsCount, GetDimensions, GetSymbolicDimensions, GetTensorShapeElementCount) to properly handle the case when a tensor does not have shape information, returning zero or empty values as appropriate.

Refactoring and Consistency

• Refactored OrtTensorTypeAndShapeInfo to store shape information in an optional structure, and updated all related methods to use pointers instead of direct objects, improving consistency and future extensibility. [1] [2]
• Updated methods for constructing shape/type info to accept pointers and handle cases where shape may be absent, including changes in onnxruntime_typeinfo.cc and related API calls. [1] [2] [3] [4] [5] [6] [7]

Graph and Proto Serialization

• Modified graph-to-proto conversion logic to propagate and respect the presence or absence of shape information, ensuring that ONNX graph serialization accurately reflects tensor shape status. [1] [2] [3] [4] [5]

These changes collectively improve the robustness and flexibility of ONNX Runtime's handling of tensor shape information, and lay the groundwork for supporting more dynamic and symbolic tensor shapes in future development.

[Copilot]

Pull Request Overview

This pull request introduces support for optional tensor shapes in ONNX Runtime, enabling the framework to handle tensors that may not have shape information. The changes refactor the tensor type and shape data structures to use optional shape storage and update all related APIs to handle cases where shape information is absent.

• Added a new API method TensorTypeAndShape_HasShape to check if a tensor has shape information
• Refactored OrtTensorTypeAndShapeInfo to store shape data in an optional structure instead of always requiring it
• Updated all shape-related methods to handle absent shape information by returning zero/empty values

Reviewed Changes

Copilot reviewed 16 out of 18 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
onnxruntime/test/testdata/abs_0d_lostdim.py	Test data generator creating ONNX model with tensors lacking shape information (shape=None)
onnxruntime/test/testdata/abs_0d_input.py	Test data generator creating ONNX model with tensors having empty shape (shape=[])
onnxruntime/test/shared_lib/test_inference.cc	Added test case OptionalShape to verify HasShape() functionality with both models
onnxruntime/test/framework/type_info_test.cc	Updated tests to use new getter methods instead of direct field access
onnxruntime/test/ep_graph/test_ep_graph.cc	Refactored test to use C++ wrapper APIs and check HasShape() before accessing shape data
onnxruntime/core/session/ort_apis.h	Added declaration for new TensorTypeAndShape_HasShape API function
onnxruntime/core/session/onnxruntime_c_api.cc	Added new API function to the OrtApi structure
onnxruntime/core/session/custom_ops.cc	Updated to use pointer-based API calls and check HasShape() before accessing shape data
onnxruntime/core/graph/graph.cc	Modified graph-to-proto conversion to handle optional shapes in tensor value info
onnxruntime/core/framework/tensor_type_and_shape.h	Major refactor introducing optional ShapeInfo structure and getter/setter methods
onnxruntime/core/framework/tensor_type_and_shape.cc	Implemented all new methods and updated existing ones to handle optional shapes
onnxruntime/core/framework/onnxruntime_typeinfo.cc	Updated factory methods to use pointer-based shape parameters
include/onnxruntime/core/session/onnxruntime_cxx_inline.h	Added implementation of HasShape() method
include/onnxruntime/core/session/onnxruntime_cxx_api.h	Added HasShape() method declaration to C++ wrapper
include/onnxruntime/core/session/onnxruntime_c_api.h	Added documentation and declaration for new HasShape API
include/onnxruntime/core/providers/utils/ort_graph_to_proto.h	Updated function signatures to include has_shape parameter and handle optional shapes

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[github-actions]

You can commit the suggested changes from lintrunner.

[snnn]

Thanks for your work on this PR! I will approve it once the linting issues are resolved.

The best fix is to remove the from onnx import TensorProto, helper import on line 10 and instead use only import onnx. All references to TensorProto and helper should then be qualified with onnx. (i.e., onnx.TensorProto, onnx.helper). This guarantees a single, clear import pattern for the file and preserves existing functionality. Specifically, in onnxruntime/test/testdata/abs_0d_lostdim.py, delete the from onnx import TensorProto, helper line, and update all usages of TensorProto and helper to use onnx.TensorProto and onnx.helper respectively (lines 19, 27, 33–50, and function definitions that use these names must be updated accordingly).