mirror of
https://github.com/saymrwulf/onnxruntime.git
synced 2026-07-27 20:02:15 +00:00
0a1902525f
### Description - Adds a patch that fixes a shape inference bug that caused a segfault: https://github.com/onnx/onnx/pull/6080 - Fix documentation describing why QLinearMatMul tests are currently being skipped. ### Motivation and Context The [PR for integrating with ONNX 1.16.0](https://github.com/microsoft/onnxruntime/pull/19745) disabled various python quantization tests due to a shape inference bug. This PR applies the ONNX fix as a patch. We still can't enable the tests because some of our CIs pip install onnx-1.16.0, which doesn't include the fix.
176 lines
7.3 KiB
Diff
176 lines
7.3 KiB
Diff
diff --git a/CMakeLists.txt b/CMakeLists.txt
|
|
index 6d7ca846..69aa622f 100644
|
|
--- a/CMakeLists.txt
|
|
+++ b/CMakeLists.txt
|
|
@@ -499,6 +499,7 @@ if (MSVC)
|
|
endif()
|
|
else()
|
|
# On non-Windows, hide all symbols we don't need
|
|
+ set(EXTRA_FLAGS "-Wno-unused-parameter")
|
|
set(ONNX_API_DEFINE "-DONNX_API=__attribute__\(\(__visibility__\(\"default\"\)\)\)")
|
|
set_target_properties(onnx_proto PROPERTIES CXX_VISIBILITY_PRESET hidden)
|
|
set_target_properties(onnx_proto PROPERTIES VISIBILITY_INLINES_HIDDEN 1)
|
|
@@ -653,20 +654,9 @@ endif()
|
|
if(MSVC)
|
|
target_compile_options(onnx_proto
|
|
PRIVATE /MP
|
|
- /wd4244 #'argument': conversion from 'google::
|
|
- #protobuf::uint64' to 'int', possible
|
|
- # loss of data
|
|
- /wd4267 # Conversion from 'size_t' to 'int',
|
|
- # possible loss of data
|
|
${EXTRA_FLAGS})
|
|
target_compile_options(onnx
|
|
PRIVATE /MP
|
|
- /wd4244 # 'argument': conversion from 'google::
|
|
- # protobuf::uint64' to 'int', possible
|
|
- # loss of data
|
|
- /wd4267 # Conversion from 'size_t' to 'int',
|
|
- # possible loss of data
|
|
- /wd4996 # The second parameter is ignored.
|
|
${EXTRA_FLAGS})
|
|
if(ONNX_USE_PROTOBUF_SHARED_LIBS)
|
|
target_compile_options(onnx_proto
|
|
diff --git a/onnx/common/file_utils.h b/onnx/common/file_utils.h
|
|
index b847798e..a6c31904 100644
|
|
--- a/onnx/common/file_utils.h
|
|
+++ b/onnx/common/file_utils.h
|
|
@@ -6,7 +6,6 @@
|
|
|
|
#pragma once
|
|
|
|
-#include <filesystem>
|
|
#include <fstream>
|
|
#include <string>
|
|
|
|
@@ -17,8 +16,7 @@ namespace ONNX_NAMESPACE {
|
|
|
|
template <typename T>
|
|
void LoadProtoFromPath(const std::string proto_path, T& proto) {
|
|
- std::filesystem::path proto_u8_path = std::filesystem::u8path(proto_path);
|
|
- std::fstream proto_stream(proto_u8_path, std::ios::in | std::ios::binary);
|
|
+ std::fstream proto_stream(proto_path, std::ios::in | std::ios::binary);
|
|
if (!proto_stream.good()) {
|
|
fail_check("Unable to open proto file: ", proto_path, ". Please check if it is a valid proto. ");
|
|
}
|
|
diff --git a/onnx/defs/quantization/defs.cc b/onnx/defs/quantization/defs.cc
|
|
index 70b4a4db..98c11545 100644
|
|
--- a/onnx/defs/quantization/defs.cc
|
|
+++ b/onnx/defs/quantization/defs.cc
|
|
@@ -200,6 +200,9 @@ ONNX_OPERATOR_SET_SCHEMA(
|
|
.SetDoc(DequantizeLinear_ver21_doc)
|
|
.TypeAndShapeInferenceFunction([](ONNX_NAMESPACE::InferenceContext& ctx) {
|
|
propagateElemTypeFromInputToOutput(ctx, 1, 0);
|
|
+ if (!hasInputShape(ctx, 0)) {
|
|
+ return;
|
|
+ }
|
|
auto& input_shape = getInputShape(ctx, 0);
|
|
updateOutputShape(ctx, 0, input_shape);
|
|
}));
|
|
diff --git a/onnx/defs/quantization/old.cc b/onnx/defs/quantization/old.cc
|
|
index 3f2d6384..d2f7cfd8 100644
|
|
--- a/onnx/defs/quantization/old.cc
|
|
+++ b/onnx/defs/quantization/old.cc
|
|
@@ -130,6 +130,9 @@ ONNX_OPERATOR_SET_SCHEMA(
|
|
.SetDoc(DequantizeLinear_ver19_doc)
|
|
.TypeAndShapeInferenceFunction([](ONNX_NAMESPACE::InferenceContext& ctx) {
|
|
propagateElemTypeFromInputToOutput(ctx, 1, 0);
|
|
+ if (!hasInputShape(ctx, 0)) {
|
|
+ return;
|
|
+ }
|
|
auto& input_shape = getInputShape(ctx, 0);
|
|
updateOutputShape(ctx, 0, input_shape);
|
|
}));
|
|
@@ -181,7 +184,6 @@ ONNX_OPERATOR_SET_SCHEMA(
|
|
if (!hasInputShape(ctx, 0)) {
|
|
return;
|
|
}
|
|
-
|
|
auto& input_shape = getInputShape(ctx, 0);
|
|
updateOutputShape(ctx, 0, input_shape);
|
|
}));
|
|
diff --git a/onnx/onnx_pb.h b/onnx/onnx_pb.h
|
|
index 0aab3e26..398ac2d6 100644
|
|
--- a/onnx/onnx_pb.h
|
|
+++ b/onnx/onnx_pb.h
|
|
@@ -47,10 +47,28 @@
|
|
#define ONNX_API ONNX_IMPORT
|
|
#endif
|
|
|
|
+#if defined(__GNUC__)
|
|
+#pragma GCC diagnostic push
|
|
+
|
|
+// In file included from onnx/onnx-ml.pb.h:30:
|
|
+// In file included from google/protobuf/extension_set.h:53:
|
|
+// google/protobuf/parse_context.h:328:47: error: implicit conversion loses integer precision: 'long' to 'int' [-Werror,-Wshorten-64-to-32]
|
|
+#if defined(__has_warning)
|
|
+#if __has_warning("-Wshorten-64-to-32")
|
|
+#pragma GCC diagnostic ignored "-Wshorten-64-to-32"
|
|
+#endif
|
|
+#endif // defined(__has_warning)
|
|
+
|
|
+#endif // defined(__GNUC__)
|
|
+
|
|
#ifdef ONNX_ML
|
|
#include "onnx/onnx-ml.pb.h"
|
|
#else
|
|
#include "onnx/onnx.pb.h"
|
|
#endif
|
|
|
|
+#if defined(__GNUC__)
|
|
+#pragma GCC diagnostic pop
|
|
+#endif
|
|
+
|
|
#endif // ! ONNX_ONNX_PB_H
|
|
diff --git a/onnx/shape_inference/implementation.cc b/onnx/shape_inference/implementation.cc
|
|
index fab1faf2..8723dcd4 100644
|
|
--- a/onnx/shape_inference/implementation.cc
|
|
+++ b/onnx/shape_inference/implementation.cc
|
|
@@ -488,29 +488,29 @@ class ShapeInferenceImplBase {
|
|
ProcessCall(n, *(iter->second), ctx);
|
|
} else {
|
|
has_unsupported_op = true;
|
|
+ return;
|
|
}
|
|
} else {
|
|
has_unsupported_op = true;
|
|
+ return;
|
|
}
|
|
- if (!has_unsupported_op) {
|
|
- for (int i = 0; i < n.output_size(); ++i) {
|
|
- // skip type and shape propagation for missing optional outputs.
|
|
- if (!n.output(i).empty())
|
|
- UpdateType(n.output(i), ctx.getOutputType(i));
|
|
- }
|
|
- // Constant values are tracked to improve inference/checking for subsequent nodes.
|
|
- ProcessConstant(n);
|
|
- // If data-propagation is enabled, partial-evaluation (aka data-propagation) is performed
|
|
- // to improve inference/checking for subsequent nodes.
|
|
- if (options.enable_data_propagation && schema && schema->has_data_propagation_function()) {
|
|
- if (generated_shape_data_by_name == nullptr) {
|
|
- fail_shape_inference(
|
|
- "Container for generated shape data cannot be nullptr when enable_data_propagation option is set.");
|
|
- }
|
|
- DataPropagationContextImpl data_propagation_ctx(
|
|
- n, value_types_by_name, input_data_by_name, *generated_shape_data_by_name);
|
|
- schema->GetDataPropagationFunction()(data_propagation_ctx);
|
|
+ for (int i = 0; i < n.output_size(); ++i) {
|
|
+ // skip type and shape propagation for missing optional outputs.
|
|
+ if (!n.output(i).empty())
|
|
+ UpdateType(n.output(i), ctx.getOutputType(i));
|
|
+ }
|
|
+ // Constant values are tracked to improve inference/checking for subsequent nodes.
|
|
+ ProcessConstant(n);
|
|
+ // If data-propagation is enabled, partial-evaluation (aka data-propagation) is performed
|
|
+ // to improve inference/checking for subsequent nodes.
|
|
+ if (options.enable_data_propagation && schema && schema->has_data_propagation_function()) {
|
|
+ if (generated_shape_data_by_name == nullptr) {
|
|
+ fail_shape_inference(
|
|
+ "Container for generated shape data cannot be nullptr when enable_data_propagation option is set.");
|
|
}
|
|
+ DataPropagationContextImpl data_propagation_ctx(
|
|
+ n, value_types_by_name, input_data_by_name, *generated_shape_data_by_name);
|
|
+ schema->GetDataPropagationFunction()(data_propagation_ctx);
|
|
}
|
|
}
|
|
ONNX_CATCH(const ONNX_NAMESPACE::InferenceError& ex) {
|