Merge pull request #789 from intel/sync_msft_25082025

Sync with Microsoft ONNX Runtime - 25/08/2025

Author: ankitm3k

.github/workflows/mac.yml

@@ -39,6 +39,7 @@ jobs:
           {"machine": "arm64", "target": "arm64", "build_config": "Debug"},
           {"machine": "arm64", "target": "arm64", "build_config": "Release"}
         ]
+
   xnnpack:
     uses: ./.github/workflows/macos-ci-build-and-test-workflow.yml
     with:
@@ -59,11 +60,13 @@ jobs:
           {"machine": "arm64", "target": "arm64", "build_config": "Debug"},
           {"machine": "arm64", "target": "arm64", "build_config": "Release"}
         ]
+
   iphone_simulator:
     runs-on: macos-15
 
     env:
       xcode_version: 16.4
+      simulator_runtime_version: 18.5
 
     strategy:
       matrix:
@@ -100,6 +103,8 @@ jobs:
             --apple_deploy_target=15.1 \
             --apple_sysroot=iphonesimulator \
             --osx_arch=${{ matrix.target_arch }}
+        env:
+          ORT_GET_SIMULATOR_DEVICE_INFO_REQUESTED_RUNTIME_VERSION: ${{ env.simulator_runtime_version }}
 
   Objective-C-StaticAnalysis:
     runs-on: macos-14

cmake/CMakeLists.txt

@@ -98,7 +98,8 @@ option(onnxruntime_USE_VSINPU "Build with VSINPU support" OFF)
 
 cmake_dependent_option(onnxruntime_USE_FLASH_ATTENTION "Build flash attention kernel for scaled dot product attention" ON "onnxruntime_USE_CUDA" OFF)
 option(onnxruntime_USE_LEAN_ATTENTION "Build lean attention kernel for scaled dot product attention" OFF)
-option(onnxruntime_USE_MEMORY_EFFICIENT_ATTENTION "Build memory efficient attention kernel for scaled dot product attention" ON)
+cmake_dependent_option(onnxruntime_USE_MEMORY_EFFICIENT_ATTENTION "Build memory efficient attention kernel for scaled dot product attention" ON "onnxruntime_USE_CUDA" OFF)
+cmake_dependent_option(onnxruntime_USE_FPA_INTB_GEMM "Build FpA IntB gemm cuda kernels" ON "onnxruntime_USE_CUDA" OFF)
 
 option(onnxruntime_BUILD_FOR_NATIVE_MACHINE "Enable this option for turning on optimization specific to this machine" OFF)
 option(onnxruntime_USE_AVX "Use AVX instructions" OFF)
@@ -696,6 +697,7 @@ if (onnxruntime_USE_CUDA)
     set(onnxruntime_USE_FLASH_ATTENTION OFF)
     set(onnxruntime_USE_LEAN_ATTENTION OFF)
     set(onnxruntime_USE_MEMORY_EFFICIENT_ATTENTION OFF)
+    set(onnxruntime_USE_FPA_INTB_GEMM OFF)
   endif()
 
   if (CMAKE_CUDA_COMPILER_VERSION VERSION_LESS 11.6)
@@ -708,6 +710,11 @@ if (onnxruntime_USE_CUDA)
     set(onnxruntime_USE_FLASH_ATTENTION OFF)
   endif()
 
+  if (CMAKE_CUDA_COMPILER_VERSION VERSION_LESS 12)
+    message( STATUS "FpA IntB Gemm unsupported for CUDA compiler version < 12.0")
+    set(onnxruntime_USE_FPA_INTB_GEMM OFF)
+  endif()
+
   if (WIN32)
     message( STATUS "Lean Attention unsupported in Windows")
     set(onnxruntime_USE_LEAN_ATTENTION OFF)
@@ -736,6 +743,11 @@ if (onnxruntime_USE_CUDA)
       message( STATUS "Enable memory efficient attention for CUDA EP")
       list(APPEND ORT_PROVIDER_FLAGS -DUSE_MEMORY_EFFICIENT_ATTENTION=1)
     endif()
+
+    if (onnxruntime_USE_FPA_INTB_GEMM)
+      message( STATUS "Enable FpA IntB Gemm for CUDA EP")
+      list(APPEND ORT_PROVIDER_FLAGS -DUSE_FPA_INTB_GEMM=1)
+    endif()
 endif()
 
 if (onnxruntime_USE_CUDA_INTERFACE AND (NOT onnxruntime_USE_CUDA))

cmake/external/onnxruntime_external_deps.cmake

@@ -331,7 +331,13 @@ if (onnxruntime_ENABLE_CPUINFO)
       set(CPUINFO_SUPPORTED TRUE)
     endif()
     if (WIN32)
-      set(CPUINFO_SUPPORTED TRUE)
+      # There's an error when linking with cpuinfo on arm64ec with a vcpkg build (--use_vcpkg).
+      # TODO Fix it and then re-enable cpuinfo on arm64ec.
+      if (onnxruntime_target_platform STREQUAL "ARM64EC")
+        set(CPUINFO_SUPPORTED FALSE)
+      else()
+        set(CPUINFO_SUPPORTED TRUE)
+      endif()
     elseif (NOT ${onnxruntime_target_platform} MATCHES "^(i[3-6]86|AMD64|x86(_64)?|armv[5-8].*|aarch64|arm64)$")
       message(WARNING
         "Target processor architecture \"${onnxruntime_target_platform}\" is not supported in cpuinfo. "

cmake/vcpkg-ports/cpuinfo/patch_cpuinfo_h_for_arm64ec.patch

@@ -0,0 +1,22 @@
+diff --git a/include/cpuinfo.h b/include/cpuinfo.h
+index f1d35d4..9e454d2 100644
+--- a/include/cpuinfo.h
++++ b/include/cpuinfo.h
+@@ -18,7 +18,7 @@
+ #define CPUINFO_ARCH_X86 1
+ #endif
+ 
+-#if defined(__x86_64__) || defined(__x86_64) || defined(_M_X64) || defined(_M_AMD64)
++#if defined(__x86_64__) || defined(__x86_64) || (defined(_M_X64) && !defined(_M_ARM64EC)) || (defined(_M_AMD64) && !defined(_M_ARM64EC))
+ #define CPUINFO_ARCH_X86_64 1
+ #endif
+ 
+@@ -26,7 +26,7 @@
+ #define CPUINFO_ARCH_ARM 1
+ #endif
+ 
+-#if defined(__aarch64__) || defined(_M_ARM64)
++#if defined(__aarch64__) || defined(_M_ARM64) || defined(_M_ARM64EC)
+ #define CPUINFO_ARCH_ARM64 1
+ #endif
+ 

cmake/vcpkg-ports/cpuinfo/portfile.cmake

@@ -9,6 +9,8 @@ vcpkg_from_github(
     REF de0ce7c7251372892e53ce9bc891750d2c9a4fd8
     SHA512 0fde9210b700d2648d37c8deeb0d5c0d007d8ca5689578dd3bce4c460886b20d7649f0194d2ea06b02238fe9d4f06193599ec3ab5cafb19f1f860b00404264fa
     HEAD_REF master
+    PATCHES
+        patch_cpuinfo_h_for_arm64ec.patch
 )
 
 vcpkg_check_features(OUT_FEATURE_OPTIONS FEATURE_OPTIONS

docs/OperatorKernels.md

@@ -703,7 +703,10 @@ Do not modify directly.*
 |GreaterOrEqual|*in* A:**T**<br> *in* B:**T**<br> *out* C:**T1**|16+|**T** = tensor(double), tensor(float), tensor(float16), tensor(int32), tensor(int64), tensor(uint32), tensor(uint64)<br/> **T1** = tensor(bool)|
 |||[12, 15]|**T** = tensor(double), tensor(float), tensor(float16), tensor(int32), tensor(int64), tensor(uint32), tensor(uint64)<br/> **T1** = tensor(bool)|
 |GridSample|*in* X:**T1**<br> *in* grid:**T2**<br> *out* Y:**T1**|16+|**T1** = tensor(float)<br/> **T2** = tensor(float)|
-|HardSigmoid|*in* X:**T**<br> *out* Y:**T**|6+|**T** = tensor(double), tensor(float), tensor(float16)|
+|HardSigmoid|*in* X:**T**<br> *out* Y:**T**|22+|**T** = tensor(bfloat16), tensor(double), tensor(float), tensor(float16)|
+|||[6, 21]|**T** = tensor(double), tensor(float), tensor(float16)|
+|HardSwish|*in* X:**T**<br> *out* Y:**T**|22+|**T** = tensor(bfloat16), tensor(double), tensor(float), tensor(float16)|
+|||[14, 21]|**T** = tensor(double), tensor(float), tensor(float16)|
 |Identity|*in* input:**T**<br> *out* output:**T**<br><br>or<br><br>*in* input:**V**<br> *out* output:**V**|19+|**V** = seq(tensor(bfloat16)), seq(tensor(bool)), seq(tensor(double)), seq(tensor(float)), seq(tensor(float16)), seq(tensor(float8e4m3fn)), seq(tensor(float8e4m3fnuz)), seq(tensor(float8e5m2)), seq(tensor(float8e5m2fnuz)), seq(tensor(int16)), seq(tensor(int32)), seq(tensor(int64)), seq(tensor(int8)), seq(tensor(uint16)), seq(tensor(uint32)), seq(tensor(uint64)), seq(tensor(uint8)), tensor(bfloat16), tensor(bool), tensor(double), tensor(float), tensor(float16), tensor(float8e4m3fn), tensor(float8e4m3fnuz), tensor(float8e5m2), tensor(float8e5m2fnuz), tensor(int16), tensor(int32), tensor(int64), tensor(int8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(uint8)|
 |||[14, 18]|**V** = seq(tensor(bfloat16)), seq(tensor(bool)), seq(tensor(double)), seq(tensor(float)), seq(tensor(float16)), seq(tensor(int16)), seq(tensor(int32)), seq(tensor(int64)), seq(tensor(int8)), seq(tensor(uint16)), seq(tensor(uint32)), seq(tensor(uint64)), seq(tensor(uint8)), tensor(bfloat16), tensor(bool), tensor(double), tensor(float), tensor(float16), tensor(int16), tensor(int32), tensor(int64), tensor(int8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(uint8)|
 |||13|**T** = tensor(bfloat16), tensor(bool), tensor(double), tensor(float), tensor(float16), tensor(int16), tensor(int32), tensor(int64), tensor(int8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(uint8)|

include/onnxruntime/core/framework/execution_provider.h

@@ -36,6 +36,7 @@ class GraphOptimizerRegistry;
 #include "core/framework/framework_provider_common.h"
 #include "core/framework/stream_handles.h"
 #include "core/framework/tuning_context.h"
+#include "core/session/onnxruntime_c_api.h"
 
 struct OrtEpDevice;
 struct OrtRunOptions;
@@ -322,6 +323,29 @@ class IExecutionProvider {
   virtual common::Status Compile(const std::vector<FusedNodeAndGraph>& fused_nodes_and_graphs,
                                  std::vector<NodeComputeInfo>& node_compute_funcs);
 
+  /**
+   * Get the compatibility info for a compiled model.
+   *
+   * The execution provider determines this value, which denotes the compatibility of the compiled model with the EP.
+   * This is stored in the model metadata under a key associated with the EP type.
+   */
+  virtual std::string GetCompiledModelCompatibilityInfo(const onnxruntime::GraphViewer& graph_viewer) const {
+    // graph_viewer and model_metadata are not used in the default implementation.
+    ORT_UNUSED_PARAMETER(graph_viewer);
+    // Default implementation returns empty string
+    return std::string();
+  }
+
+  /**
+   * Validate the compatibility of a compiled model with this execution provider.
+   */
+  virtual common::Status ValidateCompiledModelCompatibilityInfo(const std::string& /*compatibility_info*/,
+                                                                OrtCompiledModelCompatibility& model_compatibility) const {
+    // Default implementation indicates this EP does not support model compatibility validation
+    model_compatibility = OrtCompiledModelCompatibility_EP_NOT_APPLICABLE;
+    return Status::OK();
+  }
+
 #endif
 
   void SetLogger(const logging::Logger* logger) {

include/onnxruntime/core/graph/indexed_sub_graph.h

@@ -31,7 +31,7 @@ struct IndexedSubGraph {
     std::string domain;  ///< Domain of customized SubGraph/FunctionProto
     int since_version;   ///< Since version of customized SubGraph/FunctionProto.
 
-    ONNX_NAMESPACE::OperatorStatus status;  ///< Status of customized SubGraph/FunctionProto.
+    ONNX_NAMESPACE::OperatorStatus status{ONNX_NAMESPACE::OperatorStatus::STABLE};  ///< Status of customized SubGraph/FunctionProto.
 
     std::vector<std::string> inputs;                 ///< Inputs of customized SubGraph/FunctionProto.
     std::vector<std::string> outputs;                ///< Outputs of customized SubGraph/FunctionProto.

include/onnxruntime/core/providers/nv_tensorrt_rtx/nv_provider_options.h

@@ -32,9 +32,8 @@ constexpr const char* kProfilesMinShapes = "nv_profile_min_shapes";
 constexpr const char* kProfilesMaxShapes = "nv_profile_max_shapes";
 constexpr const char* kProfilesOptShapes = "nv_profile_opt_shapes";
 constexpr const char* kCudaGraphEnable = "nv_cuda_graph_enable";
-constexpr const char* kONNXBytestream = "nv_onnx_bytestream";
-constexpr const char* kONNXBytestreamSize = "nv_onnx_bytestream_size";
 constexpr const char* kMultiProfileEnable = "nv_multi_profile_enable";
+constexpr const char* kUseExternalDataInitializer = "nv_use_external_data_initializer";
 
 }  // namespace provider_option_names
 namespace run_option_names {

include/onnxruntime/core/session/onnxruntime_c_api.h

@@ -5829,7 +5829,7 @@ struct OrtApi {
    *
    * \since Version 1.23.
    */
-  ORT_API2_STATUS(Graph_GetNodes, const OrtGraph* graph,
+  ORT_API2_STATUS(Graph_GetNodes, _In_ const OrtGraph* graph,
                   _Out_writes_(num_nodes) const OrtNode** nodes, _In_ size_t num_nodes);
 
   /** \brief Get the parent node for the given graph, if any exists.
@@ -6469,6 +6469,17 @@ struct OrtApi {
                   _In_reads_(num_tensors) OrtValue* const* dst_tensors,
                   _In_opt_ OrtSyncStream* stream,
                   _In_ size_t num_tensors);
+
+  /** \brief Get ::OrtModelMetadata from an ::OrtGraph
+   *
+   * \param[in] graph The OrtGraph instance.
+   * \param[out] out Newly created ::OrtModelMetadata. Must be freed using OrtApi::ReleaseModelMetadata.
+   *
+   * \snippet{doc} snippets.dox OrtStatus Return Value
+   *
+   * \since Version 1.23.
+   */
+  ORT_API2_STATUS(Graph_GetModelMetadata, _In_ const OrtGraph* graph, _Outptr_ OrtModelMetadata** out);
 };
 
 /*