Skip to content

CVS-178009: Fix provider tests based on code review #880

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
sgbihu wants to merge 12 commits into
base: ovep-develop
Choose a base branch
from
Open

CVS-178009: Fix provider tests based on code review #880

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
12 commits
Select commit Hold shift + click to select a range
c572ba2
change the code based on review inputs
sgbihu Dec 9, 2025
8952ea3
Update the unit test based on OV test results
sgbihu Dec 11, 2025
790e552
Disable loop cases
sgbihu Dec 11, 2025
4cebc5d
The OV doesn't include the fix, need to disable again
sgbihu Dec 11, 2025
bf10bd2
Fix update comments based on review inputs
sgbihu Dec 16, 2025
8b81a24
Enable cast
sgbihu Dec 15, 2025
850f17b
Update comments
sgbihu Dec 16, 2025
71e1247
Update onnxruntime/test/providers/cpu/tensor/resize_op_test.cc
MayureshV1 Dec 16, 2025
7514048
Revert "Enable cast"
sgbihu Dec 16, 2025
ac1e9f3
Fix review inputs
sgbihu Dec 16, 2025
81ba841
Try to handle input has optimized and dynamic shape
sgbihu Dec 16, 2025
d0916b9
Add warnings
sgbihu Dec 18, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
12 changes: 12 additions & 0 deletions onnxruntime/core/providers/openvino/backends/basic_backend.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -65,6 +65,13 @@ struct OnnxToOvNetworkBindings {
[&onnx_name](const auto& ov_parameter_info) { return ov_parameter_info.get_names().contains(onnx_name); });
bool matched_names = it != ov_parameters.end();

if (it == ov_parameters.end()) {
Copy link

@ericcraw ericcraw Dec 18, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'd prefer that we only do this if onnx_input_map.size() > std::distance(ov_parameters.begin(), ov_parameters.end()) and we check it after we handle session_context.enable_causallm.

LOGS_DEFAULT(WARNING) << log_tag << "The input '" << onnx_name
<< "' is not used due to OpenVINO optimization. "
"This may cause issues if the input is required.";
continue;
}

// For Stateful Model Compilation, the ONNX model includes KV cache (past/present) tensors.
// However, these tensors are internally converted to a stateful representation, which removes them.
// It's also possible that the onnx model does not contain tensors such as beam_idx, whereas our converted
Expand Down Expand Up @@ -110,6 +117,11 @@ struct OnnxToOvNetworkBindings {

info.SetFullyDynamic(has_fully_dynamic);
info.SetBoundedDynamic(has_bounded_dynamic);
} else {
auto shape_size = ov::shape_size(shape.get_shape());
if (0 == shape_size) {
has_dynamic_io_ = true;
}
Copy link

@ericcraw ericcraw Dec 16, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can you elaborate a bit more on this case? The shape came from the OV compiled model, and the shape is reported as static but has a zero dimension? Or maybe no dimension? Does that actually make it dynamic?

Copy link
Author

@sgbihu sgbihu Dec 18, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The reason is the result is 0 size tensor. The OV can't create 0 size buffer. We need set this as dynamic to avoid allocate the buffer.
image

Copy link
Author

@sgbihu sgbihu Dec 18, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Another way is we can disable this test case due to 0 size buffer.

Copy link

@ericcraw ericcraw Dec 19, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'd prefer we don't have specific handling in i/o binding for it a case ov doesn't support, but if we keep it, just comment why it's there so it can be more easily evaluated if we want to change it in the future.

}

input_output_map.push_back(std::move(info));
Expand Down
10 changes: 9 additions & 1 deletion onnxruntime/core/providers/openvino/ov_versions/capability.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -102,8 +102,16 @@ std::vector<std::unique_ptr<ComputeCapability>> GetCapability::Execute() {
if (unsupported_nodes.empty()) {
std::vector<std::string> inputs;
std::vector<std::string> outputs;
auto input_nodes = graph_viewer_.GetInputs();
// Input is not a tensor, OV only handle a tensor input
for (auto& node : input_nodes) {
auto shape = node->Shape();
if (!shape) {
return result;
}
}
// Fill inputs with names
Iterable2String(inputs, graph_viewer_.GetInputs());
Iterable2String(inputs, input_nodes);

/* In scenarios, when there are no inputs or all inputs being initializers,
ConstantFolding optimization in onnxruntime pre-computes the value.*/
Expand Down
59 changes: 39 additions & 20 deletions onnxruntime/core/providers/openvino/ov_versions/data_ops.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -35,41 +35,22 @@ namespace openvino_ep {

// Ops which are supported only in models(as intermediate nodes) and not in unit tests
std::set<std::string> ops_supported_only_in_model = {
"Add",
"Cast",
"Celu",
"Concat",
"ConstantOfShape",
"DequantizeLinear",
"Dropout",
"Einsum",
"Exp",
"Expand",
"EyeLike",
"GatherElements",
"GatherND",
"GridSample",
"Identity",
"LayerNormalization",
"Loop",
"LSTM",
"NonMaxSuppression",
"NonZero",
"Not",
"OneHot",
"Pad",
"QuantizeLinear",
"RandomNormalLike",
"Range",
"ReduceMin",
"Resize",
"Round",
"Shape",
"Slice",
"Split",
"Tile",
"TopK",
"Trilu"};
"TopK"};

// Ops which are supported as functions (as composite ops)
std::set<std::string> ops_supported_as_function = {
Expand Down Expand Up @@ -269,6 +250,8 @@ void DataOps::populate_types_supported() {
std::make_pair(V_2020_4, ONNX_NAMESPACE::TensorProto_DataType::TensorProto_DataType_FLOAT));
supported_types_initializer_.insert(
std::make_pair(V_2020_4, ONNX_NAMESPACE::TensorProto_DataType::TensorProto_DataType_INT32));
supported_types_initializer_.insert(
std::make_pair(V_2020_4, ONNX_NAMESPACE::TensorProto_DataType::TensorProto_DataType_UINT32));
supported_types_initializer_.insert(
std::make_pair(V_2020_4, ONNX_NAMESPACE::TensorProto_DataType::TensorProto_DataType_INT64));
supported_types_initializer_.insert(
Expand Down Expand Up @@ -317,6 +300,8 @@ void DataOps::populate_types_supported() {
std::make_pair(V_2020_4, ONNX_NAMESPACE::TensorProto_DataType::TensorProto_DataType_BOOL));
supported_types_cpu_.insert(
std::make_pair(V_2020_4, ONNX_NAMESPACE::TensorProto_DataType::TensorProto_DataType_FLOAT));
supported_types_cpu_.insert(
std::make_pair(V_2020_4, ONNX_NAMESPACE::TensorProto_DataType::TensorProto_DataType_UINT32));
supported_types_cpu_.insert(
std::make_pair(V_2020_4, ONNX_NAMESPACE::TensorProto_DataType::TensorProto_DataType_INT32));
supported_types_cpu_.insert(
Expand Down Expand Up @@ -367,6 +352,7 @@ void DataOps::populate_op_mode_supported() {
no_dimension_supported_.push_back({"DynamicQuantizeLinear", V_2025_2, {"All"}});
no_dimension_supported_.push_back({"Equal", V_2022_1, {"CPU"}});
no_dimension_supported_.push_back({"Equal", V_2023_0, {"GPU"}});
no_dimension_supported_.push_back({"Exp", V_2020_4, {"CPU", "GPU"}});
no_dimension_supported_.push_back({"Expand", V_2023_3, {"CPU"}});
no_dimension_supported_.push_back({"Expand", V_2024_3, {"CPU", "GPU"}});
no_dimension_supported_.push_back({"Floor", V_2020_4, {"All"}});
Expand All @@ -382,6 +368,7 @@ void DataOps::populate_op_mode_supported() {
no_dimension_supported_.push_back({"Mul", V_2020_4, {"All"}});
no_dimension_supported_.push_back({"Neg", V_2023_0, {"CPU", "GPU"}});
no_dimension_supported_.push_back({"Pow", V_2023_0, {"CPU", "GPU"}});
no_dimension_supported_.push_back({"PRelu", V_2020_4, {"CPU", "GPU"}});
no_dimension_supported_.push_back({"QuantizeLinear", V_2021_4, {"All"}});
no_dimension_supported_.push_back({"Range", V_2021_2, {"All"}});
no_dimension_supported_.push_back({"ReduceMax", V_2021_4, {"All"}});
Expand Down Expand Up @@ -489,6 +476,38 @@ void DataOps::populate_op_mode_supported() {
}};
op_list_.insert({"Upsample", obj});
}
{
UnsupportedOpMode obj = {{V_2023_1, V_2023_2, V_2023_3, V_2024_0, V_2024_1, V_2024_2,
V_2024_3, V_2024_4, V_2024_5, V_2024_6, V_2025_0, V_2025_1, V_2025_2, V_2025_3, V_2025_4},
[this](const Node* node, const InitializedTensorSet&) {
auto& attributes = node->GetAttributes();
if (attributes.count("coordinate_transformation_mode") > 0) {
auto coordinate_transformation_mode =
attributes.at("coordinate_transformation_mode").s();
if (coordinate_transformation_mode == "tf_crop_and_resize" ||
coordinate_transformation_mode == "half_pixel_symmetric") {
return true;
}
}
if (attributes.count("antialias") > 0) {
auto antialias_mode =
attributes.at("antialias").i();
auto resize_mode = attributes.at("mode").s();
if (antialias_mode == 1 &&
(resize_mode == "linear" ||
resize_mode == "cubic")) {
return true;
}
}
if (attributes.count("exclude_outside") > 0) {
if (attributes.at("exclude_outside").i() == 1) {
return true;
}
}
return false;
}};
op_list_.insert({"Resize", obj});
}
}

bool DataOps::op_is_supported(std::string name, std::vector<SupportedOp>& op_list) {
Expand Down
74 changes: 71 additions & 3 deletions onnxruntime/test/contrib_ops/quantize_ops_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -287,9 +287,46 @@ TEST(QuantizeLinearContribOpTest, QuantizeLinear_per_tensor_float_int8) {
127, -127,
127, -128,
127, -128});
std::unordered_set<std::string> excluded_providers;
// Disable Tensorrt EP due to error: node1_quantize_scale_node: out of bounds channel axis 1. Number of input dimensions is 1.
test.Run(OpTester::ExpectResult::kExpectSuccess, "", {kTensorrtExecutionProvider});
excluded_providers.insert(kTensorrtExecutionProvider);
// Disable OV EP due to different formulation for QuantizeLinear
excluded_providers.insert(kOpenVINOExecutionProvider);
test.ConfigExcludeEps(excluded_providers)
.RunWithConfig();
}

#ifdef USE_OPENVINO
TEST(QuantizeLinearContribOpTest, OVEPQuantizeLinear_per_tensor_float_int8) {
OpTester test("QuantizeLinear", 1, onnxruntime::kMSDomain);
std::vector<int64_t> dims{16};
test.AddInput<float>("x", dims, {
0.f, 2.f, //
3.f, -3.f, // rounding half to even
2.9f, -2.9f, // low case
3.1f, -3.1f, // up case
254.f, -256.f, // critical point
255.f, -257.f, // critical point
256.f, -258.f, // critical point
1000.f, -1000.f // saturate case
});
test.AddInput<float>("y_scale", {}, {2.0f});
test.AddInput<int8_t>("y_zero_point", {}, {1});
test.AddOutput<int8_t>("y", dims,
{1, 2,
2, 0,
2, 0,
3, -1,
127, -127,
127, -128,
127, -128,
127, -128});
std::vector<std::unique_ptr<IExecutionProvider>> execution_providers;
execution_providers.emplace_back(DefaultOpenVINOExecutionProvider());
test.ConfigEps(std::move(execution_providers))
.RunWithConfig();
}
#endif // USE_OPENVINO

// Test uint16 com.microsoft.QuantizeLinear (per tensor)
TEST(QuantizeLinearContribOpTest, QuantizeLinear_per_tensor_float_uint16) {
Expand All @@ -311,10 +348,41 @@ TEST(QuantizeLinearContribOpTest, QuantizeLinear_per_tensor_float_uint16) {
32769, 32765,
65535, 0,
65535, 0});

std::unordered_set<std::string> excluded_providers;
// Disable Tensorrt EP due to error: unsupported data type
test.Run(OpTester::ExpectResult::kExpectSuccess, "", {kTensorrtExecutionProvider});
excluded_providers.insert(kTensorrtExecutionProvider);
// Disable OV EP due to different formulation for QuantizeLinear
excluded_providers.insert(kOpenVINOExecutionProvider);
test.ConfigExcludeEps(excluded_providers)
.RunWithConfig();
}

#ifdef USE_OPENVINO
TEST(QuantizeLinearContribOpTest, OVEPQuantizeLinear_per_tensor_float_uint16) {
OpTester test("QuantizeLinear", 1, onnxruntime::kMSDomain);
std::vector<int64_t> dims{12};
test.AddInput<float>("x", dims, {
0.f, -128.f, 3.f, -3.f, // rounding half to even
2.9f, -2.9f, // round < .5
3.1f, -3.1f, // round > .5
65536.f, -65534.f, // critical point
70000.f, -70000.f // saturate case
});
test.AddInput<float>("scale", {}, {2.0f}, true);
test.AddInput<uint16_t>("zero_point", {}, {32767}, true);
test.AddOutput<uint16_t>("y", dims,
{32767, 32703,
32768, 32766,
32768, 32766,
32769, 32765,
65535, 0,
65535, 0});
std::vector<std::unique_ptr<IExecutionProvider>> execution_providers;
execution_providers.emplace_back(DefaultOpenVINOExecutionProvider());
test.ConfigEps(std::move(execution_providers))
.RunWithConfig();
}
#endif // USE_OPENVINO

// Test int16 com.microsoft.QuantizeLinear (per tensor)
TEST(QuantizeLinearContribOpTest, QuantizeLinear_per_tensor_float_int16) {
Expand Down
3 changes: 2 additions & 1 deletion onnxruntime/test/providers/cpu/controlflow/loop_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1037,7 +1037,8 @@ TEST(Loop, IterationCountAsOutput) {
test.AddOutput<int64_t>("loop_var_0_final", {3, 1}, {0, 1, 2});

// Disable TensorRT on unsupported data type BOOL
test.Run(OpTester::ExpectResult::kExpectSuccess, "", {kTensorrtExecutionProvider});
// Disable OV EP due to ONNX partition create new domain and OV FE can't handle it
test.Run(OpTester::ExpectResult::kExpectSuccess, "", {kTensorrtExecutionProvider, kOpenVINOExecutionProvider});
}

#if defined(USE_CUDA)
Expand Down
5 changes: 5 additions & 0 deletions onnxruntime/test/providers/cpu/tensor/cast_op_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -75,6 +75,11 @@ void TestCastOp(gsl::span<const SrcType> input,
excluded_provider_types.insert(kCudaExecutionProvider);
}

if (input.size() == 0) {
// The OpenVINO doesn't support 0 size input
excluded_provider_types.insert(kOpenVINOExecutionProvider);
}

if (cuda_only && (excluded_provider_types.count(kCudaExecutionProvider) > 0)) {
return;
}
Expand Down
Loading
Loading