Sync with Microsoft ONNX Runtime - 31/07/2025

cmake/external/onnxruntime_external_deps.cmake

@@ -367,23 +367,23 @@ if (CPUINFO_SUPPORTED)
   set(CPUINFO_BUILD_MOCK_TESTS OFF CACHE INTERNAL "")
   set(CPUINFO_BUILD_BENCHMARKS OFF CACHE INTERNAL "")
   if (onnxruntime_target_platform STREQUAL "ARM64EC" OR onnxruntime_target_platform STREQUAL "ARM64")
-      message(STATUS "Applying a patch for Windows ARM64/ARM64EC in cpuinfo")
-      onnxruntime_fetchcontent_declare(
-        pytorch_cpuinfo
-        URL ${DEP_URL_pytorch_cpuinfo}
-        URL_HASH SHA1=${DEP_SHA1_pytorch_cpuinfo}
-        EXCLUDE_FROM_ALL
-        PATCH_COMMAND ${Patch_EXECUTABLE} -p1 < ${PROJECT_SOURCE_DIR}/patches/cpuinfo/9bb12d342fd9479679d505d93a478a6f9cd50a47.patch
-        FIND_PACKAGE_ARGS NAMES cpuinfo
-      )
+    message(STATUS "Applying a patch for Windows ARM64/ARM64EC in cpuinfo")
+    onnxruntime_fetchcontent_declare(
+      pytorch_cpuinfo
+      URL ${DEP_URL_pytorch_cpuinfo}
+      URL_HASH SHA1=${DEP_SHA1_pytorch_cpuinfo}
+      EXCLUDE_FROM_ALL
+      PATCH_COMMAND ${Patch_EXECUTABLE} -p1 < ${PROJECT_SOURCE_DIR}/patches/cpuinfo/patch_cpuinfo_h_for_arm64ec.patch
+      FIND_PACKAGE_ARGS NAMES cpuinfo
+    )
   else()
-      onnxruntime_fetchcontent_declare(
-        pytorch_cpuinfo
-        URL ${DEP_URL_pytorch_cpuinfo}
-        URL_HASH SHA1=${DEP_SHA1_pytorch_cpuinfo}
-        EXCLUDE_FROM_ALL
-        FIND_PACKAGE_ARGS NAMES cpuinfo
-      )
+    onnxruntime_fetchcontent_declare(
+      pytorch_cpuinfo
+      URL ${DEP_URL_pytorch_cpuinfo}
+      URL_HASH SHA1=${DEP_SHA1_pytorch_cpuinfo}
+      EXCLUDE_FROM_ALL
+      FIND_PACKAGE_ARGS NAMES cpuinfo
+    )
   endif()
   set(ONNXRUNTIME_CPUINFO_PROJ pytorch_cpuinfo)
   onnxruntime_fetchcontent_makeavailable(${ONNXRUNTIME_CPUINFO_PROJ})

cmake/patches/cpuinfo/9bb12d342fd9479679d505d93a478a6f9cd50a47.patch → cmake/patches/cpuinfo/patch_cpuinfo_h_for_arm64ec.patch

@@ -1,5 +1,5 @@
 diff --git a/include/cpuinfo.h b/include/cpuinfo.h
-index 6eb4b8c..4346a5a 100644
+index f1d35d4..9e454d2 100644
 --- a/include/cpuinfo.h
 +++ b/include/cpuinfo.h
 @@ -18,7 +18,7 @@
@@ -20,16 +20,3 @@ index 6eb4b8c..4346a5a 100644
  #define CPUINFO_ARCH_ARM64 1
  #endif
 
-diff --git a/src/arm/windows/init.c b/src/arm/windows/init.c
-index de2f6cc..c3a7835 100644
---- a/src/arm/windows/init.c
-+++ b/src/arm/windows/init.c
-@@ -175,7 +175,7 @@ static struct woa_chip_info* get_system_info_from_registry(void) {
- 	if (chip_info == NULL) {
- 		/* No match was found, so print a warning and assign the unknown
- 		 * case. */
--		cpuinfo_log_error(
-+		cpuinfo_log_debug(
- 			"Unknown chip model name '%ls'.\nPlease add new Windows on Arm SoC/chip support to arm/windows/init.c!",
- 			text_buffer);
- 	} else {

csharp/src/Microsoft.ML.OnnxRuntime/Exceptions.shared.cs

@@ -23,8 +23,8 @@ internal enum ErrorCode
         ModelLoaded = 8,
         NotImplemented = 9,
         InvalidGraph = 10,
-        ShapeInferenceNotRegistered = 11,
-        RequirementNotRegistered = 12,
+        ShapeInferenceNotRegistered = 11,  // TODO: should be ORT_EP_FAIL
+        RequirementNotRegistered = 12,     // TODO: should be ORT_MODEL_LOAD_CANCELED
     }
 
     /// <summary>

include/onnxruntime/core/common/status.h

@@ -46,6 +46,7 @@
   EP_FAIL = 11,
   MODEL_LOAD_CANCELED = 12,
   MODEL_REQUIRES_COMPILATION = 13,
+  NOT_FOUND = 14,
 };
 
 constexpr const char* StatusCodeToString(StatusCode status) noexcept {
@@ -78,6 +79,8 @@
       return "MODEL_LOAD_CANCELED";
     case StatusCode::MODEL_REQUIRES_COMPILATION:
       return "MODEL_REQUIRES_COMPILATION";
+    case StatusCode::NOT_FOUND:
+      return "NOT_FOUND";
     default:
       return "GENERAL ERROR";
   }
@@ -111,9 +114,11 @@
     case StatusCode::EP_FAIL:
       return HRESULT_FROM_WIN32(ERROR_INTERNAL_ERROR);
     case StatusCode::MODEL_LOAD_CANCELED:
       return HRESULT_FROM_WIN32(ERROR_CANCELLED);
     case StatusCode::MODEL_REQUIRES_COMPILATION:
       return HRESULT_FROM_WIN32(ERROR_NOT_SUPPORTED);
+    case StatusCode::NOT_FOUND:
+      return HRESULT_FROM_WIN32(ERROR_NOT_FOUND);
     default:
       return E_FAIL;
   }

include/onnxruntime/core/providers/utils/ort_graph_to_proto.h

@@ -232,7 +232,7 @@
                                                   /*out*/ std::vector<int64_t>& dims,
                                                   /*out*/ std::vector<std::string>& symbolic_dims);
 static Ort::Status OrtValueInfoToProto(const OrtValueInfo& ort_value_info, onnx::ValueInfoProto& value_info_proto);
-static Ort::Status OrtOpAttrToProto(const OrtOpAttr& ort_attr, onnx::AttributeProto& attr_proto);
+static Ort::Status OrtOpAttrToProto(const OrtNode& ort_node, const OrtOpAttr& ort_attr, onnx::AttributeProto& attr_proto);
 
 Ort::Status OrtGraphToProto(const OrtGraph& ort_graph,
                             onnx::GraphProto& graph_proto,
@@ -379,7 +379,7 @@
         }
 
         onnx::AttributeProto* attr_proto = node_proto->add_attribute();
-        ORT_EP_UTILS_CXX_RETURN_IF_ERROR(OrtOpAttrToProto(*ort_attr, *attr_proto));
+        ORT_EP_UTILS_CXX_RETURN_IF_ERROR(OrtOpAttrToProto(*ort_node, *ort_attr, *attr_proto));
       }
     }
 
@@ -652,7 +652,7 @@
   return Ort::Status{nullptr};
 }
 
-static Ort::Status OrtOpAttrToProto(const OrtOpAttr& ort_attr, onnx::AttributeProto& attr_proto) {
+static Ort::Status OrtOpAttrToProto(const OrtNode& ort_node, const OrtOpAttr& ort_attr, onnx::AttributeProto& attr_proto) {
   const OrtApi& ort_api = Ort::GetApi();
 
   const char* attr_name = nullptr;
@@ -758,6 +758,103 @@
 
       break;
     }
+    case OrtOpAttrType::ORT_OP_ATTR_TENSOR: {
+      attr_proto.set_type(onnx::AttributeProto_AttributeType_TENSOR);
+
+      onnx::TensorProto tensor_proto;
+
+      // TensorProto as an attribute value doesn't require a name.
+
+      OrtValue* ort_value = nullptr;
+      ORT_EP_UTILS_C_RETURN_IF_ERROR(ort_api.Node_GetTensorAttributeAsOrtValue(&ort_node, &ort_attr, &ort_value));
+
+      Ort::Value tensor(ort_value);
+
+      // Get tensor type and shape info
+      Ort::TensorTypeAndShapeInfo type_shape_info = tensor.GetTensorTypeAndShapeInfo();
+
+      // Get tensor type
+      ONNXTensorElementDataType element_type = type_shape_info.GetElementType();
+
+      size_t element_size = 0;
+      switch (element_type) {
+        case ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT: {
+          tensor_proto.set_data_type(onnx::TensorProto_DataType_FLOAT);
+          element_size = sizeof(float);
+          break;
+        }
+        case ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT8: {
+          tensor_proto.set_data_type(onnx::TensorProto_DataType_UINT8);
+          element_size = sizeof(uint8_t);
+          break;
+        }
+        case ONNX_TENSOR_ELEMENT_DATA_TYPE_INT8: {
+          tensor_proto.set_data_type(onnx::TensorProto_DataType_INT8);
+          element_size = sizeof(int8_t);
+          break;
+        }
+        case ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT16: {
+          tensor_proto.set_data_type(onnx::TensorProto_DataType_UINT16);
+          element_size = sizeof(uint16_t);
+          break;
+        }
+        case ONNX_TENSOR_ELEMENT_DATA_TYPE_INT16: {
+          tensor_proto.set_data_type(onnx::TensorProto_DataType_INT16);
+          element_size = sizeof(int16_t);
+          break;
+        }
+        case ONNX_TENSOR_ELEMENT_DATA_TYPE_INT32: {
+          tensor_proto.set_data_type(onnx::TensorProto_DataType_INT32);
+          element_size = sizeof(int32_t);
+          break;
+        }
+        case ONNX_TENSOR_ELEMENT_DATA_TYPE_INT64: {
+          tensor_proto.set_data_type(onnx::TensorProto_DataType_INT64);
+          element_size = sizeof(int64_t);
+          break;
+        }
+        case ONNX_TENSOR_ELEMENT_DATA_TYPE_BOOL: {
+          tensor_proto.set_data_type(onnx::TensorProto_DataType_BOOL);
+          element_size = sizeof(bool);
+          break;
+        }
+        case ONNX_TENSOR_ELEMENT_DATA_TYPE_DOUBLE: {
+          tensor_proto.set_data_type(onnx::TensorProto_DataType_DOUBLE);
+          element_size = sizeof(double);
+          break;
+        }
+        case ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT32: {
+          tensor_proto.set_data_type(onnx::TensorProto_DataType_UINT32);
+          element_size = sizeof(uint32_t);
+          break;
+        }
+        case ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT64: {
+          tensor_proto.set_data_type(onnx::TensorProto_DataType_UINT64);
+          element_size = sizeof(uint64_t);
+          break;
+        }
+        default: {
+          std::string err_msg = "Unexpected ONNXTensorElementDataType with value " + std::to_string(static_cast<int>(element_type));
+          return Ort::Status(err_msg.c_str(), ORT_FAIL);
+        }
+      }
+
+      auto shape = type_shape_info.GetShape();
+
+      for (auto& dim : shape) {
+        tensor_proto.add_dims(dim);
+      }
+
+      size_t element_count = type_shape_info.GetElementCount();
+      size_t data_bytes = element_count * element_size;
+      const void* data = tensor.GetTensorData<void>();
+
+      // Copy the Ortvalue to TensorProto as raw data
+      tensor_proto.set_raw_data(data, data_bytes);
+
+      *(attr_proto.mutable_t()) = std::move(tensor_proto);
+      break;
+    }
     default: {
       std::string err_msg = "Unexpected OrtOpAttrType with value " + std::to_string(static_cast<int>(attr_type));
       return Ort::Status(err_msg.c_str(), ORT_FAIL);

include/onnxruntime/core/session/onnxruntime_c_api.h

@@ -264,6 +264,7 @@ typedef enum OrtErrorCode {
   ORT_EP_FAIL,
   ORT_MODEL_LOAD_CANCELED,
   ORT_MODEL_REQUIRES_COMPILATION,
+  ORT_NOT_FOUND,
 } OrtErrorCode;
 
 typedef enum OrtOpAttrType {
@@ -275,6 +276,7 @@ typedef enum OrtOpAttrType {
   ORT_OP_ATTR_STRING,
   ORT_OP_ATTR_STRINGS,
   ORT_OP_ATTR_GRAPH,
+  ORT_OP_ATTR_TENSOR,
 } OrtOpAttrType;
 
 //! @}
@@ -6031,6 +6033,11 @@ struct OrtApi {
    *                           Typical usage sets this to the result of Node_GetNumAttributes(). An error status is
    *                           returned if `num_attributes` is less than the number of node attributes.
    *
+   * \note ONNX Runtime automatically sets optional (unset) attributes to their default values if the default value
+   * is a constant expression that does not depend on other tensor/model characteristics. Conv's 'kernel_shape'
+   * attribute is an example of an optional attribute that does not have a constant default value. This function
+   * does not provide any unset optional attributes without a constant default value.
+   *
    * \snippet{doc} snippets.dox OrtStatus Return Value
    *
    * \since Version 1.23.
@@ -6042,14 +6049,36 @@ struct OrtApi {
    *
    * \param[in] node The OrtNode instance.
    * \param[in] attribute_name The name of the attribute
-   * \param[out] attribute Output the attribute if its name matches 'attribute_name', otherwise output nullptr.
+   * \param[out] attribute Output parameter set to the OrtOpAttr instance if an attribute by the given name exists.
+   *                       For an unset optional attribute, `attribute` is set to NULL and a non-error status is
+   *                       returned. For an invalid attribute name, `attribute` is set to NULL and an error status with
+   *                       code ORT_NOT_FOUND is returned.
+   *
+   * \note ONNX Runtime automatically sets optional (unset) attributes to their default values if the default value
+   * is a constant expression that does not depend on other tensor/model characteristics. Conv's 'kernel_shape'
+   * attribute is an example of an optional attribute that does not have a constant default value. This function
+   * does not provide any unset optional attributes without a constant default value.
    *
    * \snippet{doc} snippets.dox OrtStatus Return Value
    *
    * \since Version 1.23.
    */
   ORT_API2_STATUS(Node_GetAttributeByName, _In_ const OrtNode* node, _In_ const char* attribute_name,
-                  _Outptr_ const OrtOpAttr** attribute);
+                  _Outptr_result_maybenull_ const OrtOpAttr** attribute);
+
+  /** \brief Get the OrtNode's 'TENSOR' attribute as an OrtValue.
+   *
+   * \param[in] node The OrtNode instance.
+   * \param[in] attribute The OrtOpAttr instance.
+   * \param[out] attr_tensor If successful, contains the 'TENSOR' attribute as a newly created OrtValue.
+                             Must be freed with OrtApi::ReleaseValue.
+   *
+   * \snippet{doc} snippets.dox OrtStatus Return Value
+   *
+   * \since Version 1.23.
+   */
+  ORT_API2_STATUS(Node_GetTensorAttributeAsOrtValue, _In_ const OrtNode* node, _In_ const OrtOpAttr* attribute,
+                  _Outptr_result_maybenull_ OrtValue** attr_tensor);
 
   /** \brief Get the attribute type as OrtOpAttrType from an OrtOpAttr.
    *

java/src/main/java/ai/onnxruntime/OrtException.java

@@ -81,11 +81,17 @@ public enum OrtErrorCode {
     /** The ONNX graph is invalid. */
     ORT_INVALID_GRAPH(10),
     /** The ORT execution provider failed. */
-    ORT_EP_FAIL(11);
+    ORT_EP_FAIL(11),
+    /** Model load was canceled. */
+    ORT_MODEL_LOAD_CANCELED(12),
+    /** Model requires compilation. */
+    ORT_MODEL_REQUIRES_COMPILATION(13),
+    /** Item was not found. */
+    ORT_NOT_FOUND(14);
 
     private final int value;
 
-    private static final OrtErrorCode[] values = new OrtErrorCode[12];
+    private static final OrtErrorCode[] values = new OrtErrorCode[15];
 
     static {
       for (OrtErrorCode ot : OrtErrorCode.values()) {

java/src/main/native/OrtJniUtil.c

@@ -1051,6 +1051,12 @@ jint convertErrorCode(OrtErrorCode code) {
             return 10;
         case ORT_EP_FAIL:
             return 11;
+        case ORT_MODEL_LOAD_CANCELED:
+            return 12;
+        case ORT_MODEL_REQUIRES_COMPILATION:
+            return 13;
+        case ORT_NOT_FOUND:
+            return 14;
         default:
             return -1; // Unknown error code
     }

onnxruntime/contrib_ops/webgpu/bert/attention_common.h

@@ -50,6 +50,7 @@ struct WebgpuAttentionParameters {
                                                                                                        v_hidden_size_(parameters.kv_hidden_size),
                                                                                                        v_head_size_(parameters.kv_hidden_size / parameters.kv_num_heads),
                                                                                                        num_heads_(parameters.num_heads),
+                                                                                                       is_unidirectional_(true),
                                                                                                        do_rotary_(parameters.do_rotary),
                                                                                                        scale_(parameters.scale),
                                                                                                        seqlen_past_kv_cache_(parameters.seqlen_past_kv_cache),