Skip to content

Add QDQ scale propagation pass #713

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

Merged
sfatimar merged 6 commits into from
Jul 3, 2025
Merged

Add QDQ scale propagation pass #713

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
05fcdce
Add pass to perform QDQ stripping and propagate scales
javier-intel Jun 15, 2025
19a5e7b
Fix disconnected outptu node
javier-intel Jun 16, 2025
66587aa
Fixes to support session.disable_quant_qdq output, remove dangling no…
javier-intel Jun 23, 2025
304a8d2
Fix lack of scales updates and remove stray QDQ nodes in certain models
javier-intel Jun 24, 2025
ce35466
Address issues with Linux CI
javier-intel Jun 25, 2025
e0cc75c
Fix for double QDQ issue
javier-intel Jul 2, 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
2 changes: 1 addition & 1 deletion cmake/onnxruntime_providers_openvino.cmake
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -49,7 +49,7 @@
endif()
add_dependencies(onnxruntime_providers_openvino onnxruntime_providers_shared ${onnxruntime_EXTERNAL_DEPENDENCIES})
target_include_directories(onnxruntime_providers_openvino SYSTEM PUBLIC ${ONNXRUNTIME_ROOT} ${CMAKE_CURRENT_BINARY_DIR} ${OpenVINO_INCLUDE_DIR} ${OPENVINO_INCLUDE_DIR_LIST} ${PYTHON_INCLUDE_DIRS} $ENV{OPENCL_INCS} $ENV{OPENCL_INCS}/../../cl_headers/)
target_link_libraries(onnxruntime_providers_openvino ${ONNXRUNTIME_PROVIDERS_SHARED} Boost::mp11 ${OPENVINO_LIB_LIST} ${ABSEIL_LIBS} Eigen3::Eigen)
target_link_libraries(onnxruntime_providers_openvino ${ONNXRUNTIME_PROVIDERS_SHARED} Boost::mp11 ${OPENVINO_LIB_LIST} ${ABSEIL_LIBS} Eigen3::Eigen onnx_proto)

target_compile_definitions(onnxruntime_providers_openvino PRIVATE FILE_NAME=\"onnxruntime_providers_openvino.dll\")

Expand Down
1 change: 1 addition & 0 deletions onnxruntime/core/optimizer/double_qdq_pairs_remover.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -52,6 +52,7 @@ static void ApplyNewInputValue(Graph& graph, Node& node, QDQ::InputIndex index,
input_init.ToProto(new_input_tensor);
auto new_name = graph.GenerateNodeArgName("DoubleQDQRemoved_" + node.InputDefs()[index]->Name());
new_input_tensor.set_name(new_name);
new_input_tensor.add_dims(1);
NodeArg& new_input = graph_utils::AddInitializerWithExternalData(graph, new_input_tensor);
graph_utils::ReplaceNodeInput(node, index, new_input);
}
Expand Down
15 changes: 13 additions & 2 deletions onnxruntime/core/providers/openvino/backend_manager.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,6 +20,7 @@
#include "core/providers/openvino/ov_interface.h"
#include "core/providers/openvino/ov_versions/capability.h"
#include "core/providers/openvino/qdq_transformations/qdq_stripping.h"
#include "core/providers/openvino/qdq_transformations/qdq_scales_fix.h"

namespace onnxruntime {
namespace openvino_ep {
Expand Down Expand Up @@ -429,8 +430,7 @@ BackendManager::GetModelProtoFromFusedNode(const onnxruntime::Node& fused_node,

const auto& onnx_model_path_name = subgraph.ModelPath();
// QDQ stripping enabled only for the NPU and experimentally on the GPU
if ((session_context_.device_type.find("NPU") != std::string::npos ||
session_context_.device_type.find("GPU") != std::string::npos) &&
if ((session_context_.device_type.find("NPU") != std::string::npos) &&
(enable_ovep_qdq_optimizer || session_context_.so_share_ep_contexts)) {
std::unique_ptr<onnxruntime::Model> model;
Status status = CreateModelWithStrippedQDQNodes(subgraph, logger, session_context_.so_share_ep_contexts, enable_ovep_qdq_optimizer, model, shared_context_.shared_weights);
Expand All @@ -440,6 +440,17 @@ BackendManager::GetModelProtoFromFusedNode(const onnxruntime::Node& fused_node,
DumpOpenVINOEPModel(onnx_model_path_name, model_proto.get(), fused_node);
ORT_ENFORCE(status.IsOK(), status.ErrorMessage());
return model_proto;
} else if ((session_context_.device_type.find("GPU") != std::string::npos) &&
enable_ovep_qdq_optimizer) {
// Create a copy of the model
std::unique_ptr<onnxruntime::Model> model;
Status status = qdq_scales_fix::Transform(subgraph, logger, model);
Copy link

@preetha-intel preetha-intel Jun 18, 2025

Choose a reason for hiding this comment

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

Is this pass happening even for non quantized models?

Copy link

@mklimenk mklimenk Jul 3, 2025 •
edited
Loading

Choose a reason for hiding this comment

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

@preetha-intel, this pass is happening only when the enable_qdq_optimizer flag is set.
Inside the pass it specifically looks for quantized blocks with (u)int16 precision and ignores everything else. So the regular models are not affected by it, even if the flag was passed by accident

auto model_proto = model->ToProto();
model_proto->set_ir_version(ONNX_NAMESPACE::Version::IR_VERSION);
print_model_proto_duration();
DumpOpenVINOEPModel(onnx_model_path_name, model_proto.get(), fused_node);
ORT_ENFORCE(status.IsOK(), status.ErrorMessage());
return model_proto;
} else {
LOGS_DEFAULT(INFO) << "[OpenVINO-EP] OVEP QDQ optimization pass is disabled";
auto model = subgraph.CreateModel(logger);
Expand Down
24 changes: 24 additions & 0 deletions onnxruntime/core/providers/openvino/ov_protobuf_utils.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,24 @@
// Copyright (C) Intel Corporation
// Licensed under the MIT License

#include "ov_protobuf_utils.h"

Check notice on line 4 in onnxruntime/core/providers/openvino/ov_protobuf_utils.cpp

View workflow job for this annotation

GitHub Actions / cpplint

[cpplint] onnxruntime/core/providers/openvino/ov_protobuf_utils.cpp#L4 

 Include the directory when naming header files  [build/include_subdir] [4]
Raw output 
onnxruntime/core/providers/openvino/ov_protobuf_utils.cpp:4:  Include the directory when naming header files  [build/include_subdir] [4]

#include "core/graph/onnx_protobuf.h"
#include "core/common/common.h"

namespace onnxruntime {
namespace openvino_ep {
float get_float_initializer_data(const void* initializer) {
const auto* tp = reinterpret_cast<const ONNX_NAMESPACE::TensorProto*>(initializer);
ORT_ENFORCE((tp->has_data_type() && (tp->data_type() == ONNX_NAMESPACE::TensorProto_DataType_FLOAT)));
// ORT_ENFORCE(initializer.dims_size() == 1);
return tp->float_data(0);
}
void set_float_initializer_data(const void* initializer, float data) {
auto* tp = (ONNX_NAMESPACE::TensorProto*)(initializer);
ORT_ENFORCE((tp->has_data_type() && (tp->data_type() == ONNX_NAMESPACE::TensorProto_DataType_FLOAT)));
// ORT_ENFORCE(initializer.dims_size() == 1);
tp->set_float_data(0, data);
}
} // namespace openvino_ep
} // namespace onnxruntime
10 changes: 10 additions & 0 deletions onnxruntime/core/providers/openvino/ov_protobuf_utils.h
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,10 @@
// Copyright (C) Intel Corporation
// Licensed under the MIT License

#pragma once
namespace onnxruntime {
namespace openvino_ep {
float get_float_initializer_data(const void* initializer);
void set_float_initializer_data(const void* initializer, float data);
}
} // namespace onnxruntime
Loading
Loading