Fix CoreML reshape to support dynamic shapes

onnxruntime/core/providers/coreml/builders/impl/reshape_op_builder.cc

@@ -40,15 +40,19 @@ Status ReshapeOpBuilder::AddToModelBuilderImpl(ModelBuilder& model_builder,
                                                const logging::Logger& logger) const {
   const auto& input_defs = node.InputDefs();
   std::vector<int64_t> input_shape;
-  ORT_RETURN_IF_NOT(GetStaticShape(*input_defs[0], input_shape, logger), "Cannot get shape of data");
+  ORT_RETURN_IF_NOT(GetShape(*input_defs[0], input_shape, logger), "Cannot get shape of data");
 
   const auto& data_name = input_defs[0]->Name();
   const auto& new_shape_name = input_defs[1]->Name();
   Initializer unpacked_tensor(model_builder.GetGraphViewer().GetGraph(), *model_builder.GetConstantInitializer(new_shape_name));
   TensorShapeVector new_shape = ToShapeVector(unpacked_tensor.DataAsSpan<int64_t>());
 
-  // ReshapeHelper applies the ONNX rules to create the concrete output shape
-  ReshapeHelper helper(TensorShape(input_shape), new_shape);
+  // ReshapeHelper applies the ONNX rules to create the concrete output shape.
+  // Only use it if the input shape is static (no dynamic dimensions).
+  // CoreML MIL supports -1 in the shape and can infer at runtime.
+  if (IsStaticShape(input_shape)) {
+    ReshapeHelper helper(TensorShape(input_shape), new_shape);
+  }
 
   if (model_builder.CreateMLProgram()) {
     using namespace CoreML::Specification::MILSpec;
@@ -96,7 +100,7 @@ bool ReshapeOpBuilder::IsOpSupportedImpl(const Node& node,
   }
 
   std::vector<int64_t> input_shape;
-  if (!GetStaticShape(*input_defs[0], input_shape, logger))
+  if (!GetShape(*input_defs[0], input_shape, logger))
     return false;
 
   if (input_shape.empty()) {

onnxruntime/test/providers/coreml/dynamic_input_test.cc

@@ -84,6 +84,41 @@ TEST(CoreMLExecutionProviderDynamicInputShapeTest, MobileNetExcerpt) {
   }
 }
 
+TEST(CoreMLExecutionProviderDynamicInputShapeTest, Reshape) {
+  constexpr auto model_path = ORT_TSTR("testdata/reshape_with_dynamic_input_shape.onnx");
+
+  auto test = [&](const size_t M) {
+    SCOPED_TRACE(MakeString("M=", M));
+    std::unordered_map<std::string, std::string> options;
+    auto coreml_ep = CoreMLProviderFactoryCreator::Create(options)->CreateProvider();
+
+    const auto ep_verification_params = EPVerificationParams{
+        ExpectedEPNodeAssignment::All,
+        1e-4f,
+    };
+
+#if defined(__APPLE__)
+    RandomValueGenerator gen{1234};
+    // Input is [M, 512], reshape to [-1, 2048] so M must be a multiple of 4.
+    const auto X_shape = std::vector<int64_t>{static_cast<int64_t>(M * 4), 512};
+    const auto X_data = gen.Uniform<float>(X_shape, 0.0f, 1.0f);
+
+    OrtValue X = CreateInputOrtValueOnCPU<float>(X_shape, X_data);
+
+    RunAndVerifyOutputsWithEP(model_path, CurrentTestName(),
+                              std::move(coreml_ep),
+                              {{"X", X}},
+                              ep_verification_params);
+#else
+    TestModelLoad(model_path, std::move(coreml_ep), ep_verification_params.ep_node_assignment);
+#endif
+  };
+
+  for (size_t i = 1; i <= 3; ++i) {
+    test(i);
+  }
+}
+
 TEST(CoreMLExecutionProviderDynamicInputShapeTest, EmptyInputFails) {
   constexpr auto model_path = ORT_TSTR("testdata/matmul_with_dynamic_input_shape.onnx");
 

onnxruntime/test/testdata/reshape_with_dynamic_input_shape.py

@@ -0,0 +1,33 @@
+from pathlib import Path
+
+import onnx
+from onnx import TensorProto, helper
+
+# This model contains a Reshape where:
+# - X has shape [M, 512] and `M` is a dynamic dimension.
+# - shape is a constant initializer with value [-1, 2048].
+# CoreML MIL supports -1 in the shape and can infer the dimension at runtime.
+
+M = "M"
+K = 512
+N = 2048
+
+graph = helper.make_graph(
+    [  # nodes
+        helper.make_node("Reshape", ["X", "shape"], ["Y"], "Reshape"),
+    ],
+    "ReshapeWithDynamicInputShape",  # name
+    [  # inputs
+        helper.make_tensor_value_info("X", TensorProto.FLOAT, [M, K]),
+    ],
+    [  # outputs
+        helper.make_tensor_value_info("Y", TensorProto.FLOAT, [None, N]),
+    ],
+    [  # initializers
+        helper.make_tensor("shape", TensorProto.INT64, [2], [-1, N]),
+    ],
+)
+
+opset_imports = [helper.make_operatorsetid("", 19)]
+model = helper.make_model(graph, opset_imports=opset_imports)
+onnx.save(model, str(Path(__file__).parent / "reshape_with_dynamic_input_shape.onnx"))