Fix ModelEditorValueInfoToOnnx dropping symbolic dim names

[f-dy]

Description

ModelEditorValueInfoToOnnx in onnxruntime/core/graph/graph.cc had a missing increment bug: size_t idx = 0; was declared before the dim loop but never incremented, so dim_params[idx] always read the first entry. Every dynamic dimension in the generated ValueInfoProto was assigned the same name — the name of the first dim — instead of its own.

For a ValueInfo built with shape [-1, -1, -1] and symbolic dim names ["N", "H", "W"], the resulting session exposes dims ["N", "N", "N"] via SessionGetInputTypeInfo → GetSymbolicDimensions.

Why this wasn't caught earlier

The only existing SetSymbolicDimensions coverage is indirect (ModelEditorAPITest.Basic_CxxApi), and it only passes one non-empty symbolic name ("multiple_of_3" on dim 0, empty on dim 1). With exactly one name, the bug's "always read dim_params[0]" behavior happens to produce the correct result, so the test passes despite the bug.

The bug is only observable with ≥2 distinct non-empty symbolic names.

Notably, the comment in ModelEditorAPITest.Basic_CApi already documents the expected behavior:

"call SetDimensions with {-1, 3, 2} and SetSymbolicDimensions with {"N", nullptr, nullptr} to create a shape of {"N", 3, 2}"

The comment is correct; the implementation just didn't match.

Impact

Downstream, this causes the TensorRT EP to reject valid dynamic shape profiles. When trt_profile_{min,opt,max}_shapes bind distinct ranges to each dim, TRT (correctly) sees all ranges attached to the same symbolic name and errors with:

N: Input dimensions with this name have different constant values
or have contradictory IOptimizationProfile kMIN/kMAX constraints.
In getBounds at /_src/optimizer/common/shape/shapeContext.cpp:4516

I tripped over this in a local project that uses the Model Editor API to bake letterbox preprocessing onto a pose-estimation ONNX model. Python works fine (Python builds the ModelProto directly via onnx.helper and bypasses the Model Editor API); the equivalent C++ code hits this issue.

The fix

Index dim_params by the same position as dims:

     if (tensor_info.HasShape()) {
       auto& shape = *tensor->mutable_shape();

-      size_t idx = 0;
       const auto dims = tensor_info.GetShape()->GetDims();
       const auto& dim_params = *tensor_info.GetDimParams();
-      for (auto dim : dims) {
+      for (size_t idx = 0; idx < dims.size(); ++idx) {
+        const auto dim = dims[idx];
         auto& dim_proto = *shape.add_dim();
         if (dim >= 0) {
           dim_proto.set_dim_value(dim);
         } else {
           const std::string& dim_param = dim_params[idx];
           // if empty leave the new dim_proto with neither dim_value nor dim_param set. this represents an 'unknown' dim
           if (!dim_param.empty()) {
             dim_proto.set_dim_param(dim_param);
           }
         }
       }
     }

This restores the invariant that dim_params[i] corresponds to dims[i], which is already how:

• OrtApis::SetSymbolicDimensions stores names (tensor_type_and_shape.cc)
• OrtApis::GetSymbolicDimensions reads names (tensor_type_and_shape.cc)
• OrtTypeInfo::FromTypeProto reconstructs the info in the reverse direction (onnxruntime_typeinfo.cc)

Regression tests

Two new tests in onnxruntime/test/shared_lib/test_model_builder_api.cc:

SymbolicDimensions_DistinctNamesPreserved — builds a minimal Identity model via the C++ Model Editor API with three named dynamic dims {"N", "H", "W"}, finalizes the session, and asserts GetInputTypeInfo(0) → GetSymbolicDimensions() returns {"N", "H", "W"}. Without the fix, the second and third names would both be "N" (the first name).

SymbolicDimensions_PartiallyNamedDimensions — covers the mixed case where some dims are named, some are static, and some are dynamic-but-unnamed. Shape [-1, 3, -1] with names {"N", "", ""}. This matches the exact example given in the existing Basic_CApi test's doc comment.

The bug was first observed in a separate C++ application using the Model Editor API to bake preprocessing onto a YOLO-pose model; GetInputTypeInfo after FinalizeModelEditorSession consistently returned ['N','N','N',3] instead of the ['N','H','W',3] that was set via SetSymbolicDimensions. These regression tests exercise the same code path within ORT's own test suite so CI will catch any re-introduction.

Motivation and context

This blocks using the C++ Model Editor API to build models with multi-named dynamic shapes that must then run through the TensorRT EP with dynamic shape profiles — a reasonable use case for any application that bakes preprocessing on top of a dynamic-input model (e.g. letterboxing for detection / pose networks).

[f-dy]

@microsoft-github-policy-service agree company="Amazon"