Merge pull request #989 from intel/sync_msft_25032026

Sync with Microsoft ONNX Runtime - 25032026

Author: Jaswanth51

.github/workflows/windows_webgpu.yml

@@ -155,6 +155,86 @@ jobs:
         working-directory: ${{ github.workspace }}\csharp
         continue-on-error: true
 
+  webgpu_plugin_build_x64_RelWithDebInfo:
+    runs-on: [
+      "self-hosted",
+      "1ES.Pool=onnxruntime-github-Win2022-GPU-A10",
+      "JobId=webgpu_plugin_build_x64_RelWithDebInfo-${{ github.run_id }}-${{ github.run_number }}-${{ github.run_attempt }}"
+      ]
+    timeout-minutes: 300
+    env:
+      OnnxRuntimeBuildDirectory: ${{ github.workspace }}
+      setVcvars: true
+      ALLOW_RELEASED_ONNX_OPSET_ONLY: "0"
+      DocUpdateNeeded: false
+      NVIDIA_TF32_OVERRIDE: "0"
+      ONNXRUNTIME_TEST_GPU_DEVICE_ID: "0"
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+          submodules: none
+
+      - name: Setup Python 3.12
+        uses: actions/setup-python@v6
+        with:
+          python-version: "3.12"
+          architecture: x64
+
+      - name: Locate vcvarsall and Setup Env
+        uses: ./.github/actions/locate-vcvarsall-and-setup-env
+        with:
+          architecture: x64
+
+      - name: Install python modules
+        run: python -m pip install -r tools\ci_build\github\windows\python\requirements.txt
+        shell: cmd
+        working-directory: ${{ github.workspace }}
+
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
+        with:
+          node-version: "20.x"
+
+      - uses: actions/cache@v5
+        id: onnx-node-tests-cache
+        with:
+          path: ${{ github.workspace }}/js/test/
+          key: onnxnodetests-${{ hashFiles('js/scripts/prepare-onnx-node-tests.ts') }}
+
+      - name: Build and Test
+        shell: pwsh
+        run: |
+          python.exe ${{ github.workspace }}\tools\ci_build\build.py `
+            --config RelWithDebInfo `
+            --build_dir ${{ github.workspace }} `
+            --skip_submodule_sync `
+            --parallel `
+            --use_binskim_compliant_compile_flags `
+            --cmake_generator "Visual Studio 17 2022" `
+            --enable_onnx_tests `
+            --use_webgpu shared_lib `
+            --wgsl_template static `
+            --use_vcpkg --use_vcpkg_ms_internal_asset_cache `
+            --cmake_extra_defines onnxruntime_BUILD_UNIT_TESTS=ON onnxruntime_ENABLE_DAWN_BACKEND_D3D12=1 onnxruntime_ENABLE_DAWN_BACKEND_VULKAN=1 `
+            --disable_rtti `
+            --enable_lto
+
+          if ($lastExitCode -ne 0) {
+            exit $lastExitCode
+          }
+
+      - name: Publish artifacts
+        uses: actions/upload-artifact@v4
+        with:
+          name: webgpu-plugin-binaries
+          path: |
+            ${{ github.workspace }}/RelWithDebInfo/RelWithDebInfo/onnxruntime_providers_webgpu.dll
+            ${{ github.workspace }}/RelWithDebInfo/RelWithDebInfo/onnxruntime_providers_webgpu.pdb
+            ${{ github.workspace }}/RelWithDebInfo/RelWithDebInfo/dxcompiler.dll
+            ${{ github.workspace }}/RelWithDebInfo/RelWithDebInfo/dxil.dll
+
   webgpu_external_dawn_build_x64_RelWithDebInfo:
     runs-on: [
         "self-hosted",

cmake/onnxruntime_mlas.cmake

@@ -34,6 +34,8 @@ onnxruntime_add_static_library(onnxruntime_mlas
   ${MLAS_SRC_DIR}/eltwise.h
   ${MLAS_SRC_DIR}/eltwise.cpp
   ${MLAS_SRC_DIR}/erf.cpp
+  ${MLAS_SRC_DIR}/silu.cpp
+  ${MLAS_SRC_DIR}/gelu.cpp
   ${MLAS_SRC_DIR}/compute.cpp
   ${MLAS_SRC_DIR}/dequantize.cpp
   ${MLAS_SRC_DIR}/quantize.cpp
@@ -201,6 +203,14 @@ function(setup_mlas_source_for_windows)
     )
     set_source_files_properties(${mlas_platform_srcs_avx2} PROPERTIES COMPILE_FLAGS "/arch:AVX2")
 
+    set(mlas_platform_srcs_avx512
+      ${MLAS_SRC_DIR}/intrinsics/avx512/gelu_avx512f.cpp
+      ${MLAS_SRC_DIR}/intrinsics/avx512/silu_avx512f.cpp
+      ${MLAS_SRC_DIR}/intrinsics/avx512/quantize_avx512f.cpp
+    )
+
+    set_source_files_properties(${mlas_platform_srcs_avx512} PROPERTIES COMPILE_FLAGS "/arch:AVX512")
+
     target_sources(onnxruntime_mlas PRIVATE
       ${MLAS_SRC_DIR}/dgemm.cpp
       ${mlas_platform_srcs_avx}
@@ -212,7 +222,7 @@ function(setup_mlas_source_for_windows)
       ${MLAS_SRC_DIR}/qgemm_kernel_avx2.cpp
       ${MLAS_SRC_DIR}/qgemm_kernel_sse.cpp
       ${MLAS_SRC_DIR}/qgemm_kernel_sse41.cpp
-      ${MLAS_SRC_DIR}/intrinsics/avx512/quantize_avx512f.cpp
+      ${mlas_platform_srcs_avx512}
       ${MLAS_SRC_DIR}/sqnbitgemm_lut_kernel_avx2.h
       ${MLAS_SRC_DIR}/sqnbitgemm_lut_kernel_avx2.cpp
       ${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx2.cpp
@@ -764,6 +774,8 @@ endif()
           ${MLAS_SRC_DIR}/x86_64/SoftmaxKernelAvx512F.S
           ${MLAS_SRC_DIR}/x86_64/SpoolKernelAvx512F.S
           ${MLAS_SRC_DIR}/x86_64/TransKernelAvx512F.S
+          ${MLAS_SRC_DIR}/intrinsics/avx512/gelu_avx512f.cpp
+          ${MLAS_SRC_DIR}/intrinsics/avx512/silu_avx512f.cpp
           ${MLAS_SRC_DIR}/intrinsics/avx512/quantize_avx512f.cpp
         )
         set_source_files_properties(${mlas_platform_srcs_avx512f} PROPERTIES COMPILE_FLAGS "-mavx512f")

cmake/onnxruntime_providers_webgpu.cmake

@@ -56,7 +56,7 @@
     endif()
     source_group(TREE ${ONNXRUNTIME_ROOT} FILES ${onnxruntime_providers_webgpu_cc_srcs})
 
-    onnxruntime_add_shared_library(onnxruntime_providers_webgpu ${onnxruntime_providers_webgpu_cc_srcs})
+    onnxruntime_add_shared_library_module(onnxruntime_providers_webgpu ${onnxruntime_providers_webgpu_cc_srcs})
     onnxruntime_add_include_to_target(onnxruntime_providers_webgpu
         ${REPO_ROOT}/include/onnxruntime/core/session
         onnxruntime_common
@@ -119,6 +119,12 @@
     if (CMAKE_SYSTEM_NAME STREQUAL "Emscripten")
       message(FATAL_ERROR "WebGPU EP shared library build is not supported on Emscripten. Please use static library build.")
     endif()
+
+    # Configure precompiled headers for shared library build
+    # PCH ensures ep/adapters.h is included first and improves compilation speed
+    target_precompile_headers(onnxruntime_providers_webgpu PRIVATE
+      "${REPO_ROOT}/include/onnxruntime/ep/adapters.h"
+    )
   endif()
 
   set_target_properties(onnxruntime_providers_webgpu PROPERTIES CXX_STANDARD_REQUIRED ON)

cmake/onnxruntime_unittests.cmake

@@ -1042,6 +1042,18 @@ function(onnxruntime_apply_test_target_workarounds target)
   endif()
 endfunction()
 
+# Set environment variables for plugin EP tests when run via CTest.
+function(onnxruntime_set_plugin_ep_test_environment target)
+  if(onnxruntime_USE_WEBGPU AND onnxruntime_USE_EP_API_ADAPTERS)
+    set(ORT_PLUGIN_EP_JSON_CONFIG "{\"ep_library_registration_name\": \"WebGPU_PluginEP\", \"ep_library_path\": \"$<TARGET_FILE_NAME:onnxruntime_providers_webgpu>\", \"selected_ep_name\": \"WebGpuExecutionProvider\"}")
+    set_tests_properties(${target} PROPERTIES
+      ENVIRONMENT "ORT_UNIT_TEST_MAIN_DYNAMIC_PLUGIN_EP_CONFIG_JSON=${ORT_PLUGIN_EP_JSON_CONFIG}"
+    )
+  # TODO: add for other plugin EPs if needed
+  # elseif()
+  endif()
+endfunction()
+
 function(onnxruntime_apply_emscripten_test_link_settings target)
   if (CMAKE_SYSTEM_NAME STREQUAL "Emscripten")
     set_target_properties(${target} PROPERTIES LINK_DEPENDS ${TEST_SRC_DIR}/wasm/onnxruntime_test_adapter.js)
@@ -1250,6 +1262,7 @@ block()
   )
 
   onnxruntime_apply_test_target_workarounds(onnxruntime_provider_test)
+  onnxruntime_set_plugin_ep_test_environment(onnxruntime_provider_test)
 
   # Expose QNN SDK headers to unit tests via an interface target
   if(onnxruntime_USE_QNN)

docs/OperatorKernels.md

@@ -103,6 +103,8 @@ Do not modify directly.*
 |||[11, 13]|**T** = tensor(double), tensor(float), tensor(int32), tensor(int64)<br/> **T2** = tensor(int32), tensor(int64)|
 |DFT|*in* input:**T1**<br> *in* dft_length:**T2**<br> *in* axis:**tensor(int64)**<br> *out* output:**T1**<br><br>or<br><br>*in* input:**T1**<br> *in* dft_length:**T2**<br> *out* output:**T1**|20+|**T1** = tensor(double), tensor(float)<br/> **T2** = tensor(int32), tensor(int64)|
 |||[17, 19]|**T1** = tensor(double), tensor(float)<br/> **T2** = tensor(int32), tensor(int64)|
+|DeformConv|*in* X:**T**<br> *in* W:**T**<br> *in* offset:**T**<br> *in* B:**T**<br> *in* mask:**T**<br> *out* Y:**T**|22+|**T** = tensor(double), tensor(float)|
+|||[19, 21]|**T** = tensor(double), tensor(float)|
 |DepthToSpace|*in* input:**T**<br> *out* output:**T**|13+|**T** = tensor(double), tensor(float), tensor(uint8)|
 |||[11, 12]|**T** = tensor(double), tensor(float), tensor(uint8)|
 |||[1, 10]|**T** = tensor(double), tensor(float)|
@@ -697,6 +699,8 @@ Do not modify directly.*
 |Crop|*in* input:**T**<br> *out* output:**T**|1+|**T** = tensor(double), tensor(float), tensor(float16)|
 |CumSum|*in* x:**T**<br> *in* axis:**T2**<br> *out* y:**T**|14+|**T** = tensor(double), tensor(float), tensor(float16), tensor(int32), tensor(int64), tensor(uint32), tensor(uint64)<br/> **T2** = tensor(int32), tensor(int64)|
 |||[11, 13]|**T** = tensor(double), tensor(float), tensor(int32), tensor(int64), tensor(uint32), tensor(uint64)<br/> **T2** = tensor(int32), tensor(int64)|
+|DeformConv|*in* X:**T**<br> *in* W:**T**<br> *in* offset:**T**<br> *in* B:**T**<br> *in* mask:**T**<br> *out* Y:**T**|22+|**T** = tensor(bfloat16), tensor(double), tensor(float), tensor(float16)|
+|||[19, 21]|**T** = tensor(double), tensor(float), tensor(float16)|
 |DepthToSpace|*in* input:**T**<br> *out* output:**T**|13+|**T** = tensor(double), tensor(float), tensor(float16)|
 |||[11, 12]|**T** = tensor(double), tensor(float), tensor(float16)|
 |||[1, 10]|**T** = tensor(double), tensor(float), tensor(float16)|
@@ -843,7 +847,12 @@ Do not modify directly.*
 |PRelu|*in* X:**T**<br> *in* slope:**T**<br> *out* Y:**T**|16+|**T** = tensor(double), tensor(float), tensor(float16)|
 |||[9, 15]|**T** = tensor(double), tensor(float), tensor(float16)|
 |||[7, 8]|**T** = tensor(double), tensor(float), tensor(float16)|
-|Pad|*in* data:**T**<br> *in* pads:**tensor(int64)**<br> *in* constant_value:**T**<br> *in* axes:**Tind**<br> *out* output:**T**<br><br>or<br><br>*in* data:**T**<br> *in* pads:**tensor(int64)**<br> *in* constant_value:**T**<br> *out* output:**T**<br><br>or<br><br>*in* data:**T**<br> *out* output:**T**|18+|**T** = tensor(bool), tensor(double), tensor(float), tensor(float16)|
+|Pad|*in* data:**T**<br> *in* pads:**tensor(int64)**<br> *in* constant_value:**T**<br> *in* axes:**Tind**<br> *out* output:**T**<br><br>or<br><br>*in* data:**T**<br> *in* pads:**tensor(int64)**<br> *in* constant_value:**T**<br> *out* output:**T**<br><br>or<br><br>*in* data:**T**<br> *out* output:**T**|25+|**T** = tensor(bool), tensor(double), tensor(float), tensor(float16)|
+|||24|**T** = tensor(bool), tensor(double), tensor(float), tensor(float16)|
+|||23|**T** = tensor(bool), tensor(double), tensor(float), tensor(float16)|
+|||[21, 22]|**T** = tensor(bool), tensor(double), tensor(float), tensor(float16)|
+|||[19, 20]|**T** = tensor(bool), tensor(double), tensor(float), tensor(float16)|
+|||18|**T** = tensor(bool), tensor(double), tensor(float), tensor(float16)|
 |||[13, 17]|**T** = tensor(bool), tensor(double), tensor(float), tensor(float16)|
 |||[11, 12]|**T** = tensor(double), tensor(float), tensor(float16)|
 |||[2, 10]|**T** = tensor(double), tensor(float), tensor(float16)|
@@ -902,7 +911,9 @@ Do not modify directly.*
 |||[11, 12]|**T1** = tensor(double), tensor(float), tensor(float16), tensor(int32), tensor(uint8)|
 |||10|**T** = tensor(double), tensor(float), tensor(float16), tensor(int32), tensor(uint8)|
 |ReverseSequence|*in* input:**T**<br> *in* sequence_lens:**tensor(int64)**<br> *out* Y:**T**|10+|**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)|
-|RoiAlign|*in* X:**T1**<br> *in* rois:**T1**<br> *in* batch_indices:**T2**<br> *out* Y:**T1**|10+|**T1** = tensor(double), tensor(float)<br/> **T2** = tensor(int64)|
+|RoiAlign|*in* X:**T1**<br> *in* rois:**T1**<br> *in* batch_indices:**T2**<br> *out* Y:**T1**|22+|**T1** = tensor(bfloat16), tensor(double), tensor(float), tensor(float16)<br/> **T2** = tensor(int64)|
+|||[16, 21]|**T1** = tensor(double), tensor(float), tensor(float16)<br/> **T2** = tensor(int64)|
+|||[10, 15]|**T1** = tensor(double), tensor(float)<br/> **T2** = tensor(int64)|
 |RotaryEmbedding|*in* X:**T**<br> *in* cos_cache:**T**<br> *in* sin_cache:**T**<br> *in* position_ids:**M**<br> *out* Y:**T**|23+|**M** = tensor(int64)<br/> **T** = tensor(bfloat16), tensor(float), tensor(float16)|
 |Round|*in* X:**T**<br> *out* Y:**T**|11+|**T** = tensor(double), tensor(float), tensor(float16)|
 |ScaledTanh|*in* input:**T**<br> *out* output:**T**|1+|**T** = tensor(double), tensor(float), tensor(float16)|

include/onnxruntime/core/session/onnxruntime_session_options_config_keys.h

@@ -391,6 +391,13 @@ static const char* const kOrtSessionOptionsMlasDisableKleidiAi = "mlas.disable_k
 // If not provided, default is 4.
 static const char* const kOrtSessionOptionsQDQMatMulNBitsAccuracyLevel = "session.qdq_matmulnbits_accuracy_level";
 
+// Block size used when converting per-tensor or per-axis DQ + MatMul to MatMulNBits.
+// Only applies to DQ nodes without an existing block_size attribute (i.e., per-tensor or per-axis quantization).
+// Positive value: explicit block_size (must be power-of-2 and >= 16, e.g., 16, 32, 64, 128).
+// "0" or not provided: use default block_size of 32.
+// "-1": heuristic - largest power-of-2 <= min(K, 256) that minimizes padding.
+static const char* const kOrtSessionOptionsQDQMatMulNBitsBlockSize = "session.qdq_matmulnbits_block_size";
+
 // Enable the DQ->MatMulNBits fusion graph transformer.
 // "0": disabled (default). "1": enabled.
 // This is typically set automatically by InferenceSession when the NvTensorRTRTX EP is registered.

include/onnxruntime/ep/adapter/allocator.h

@@ -18,23 +18,50 @@ namespace adapter {
 /// </summary>
 class Allocator : public OrtAllocator {
  public:
+  /**
+   * Create from an existing AllocatorPtr.
+   */
   explicit Allocator(const OrtMemoryInfo* memory_info, AllocatorPtr impl)
-      : OrtAllocator{}, memory_info_(memory_info), impl_(impl) {
+      : Allocator{memory_info} {
+    ORT_ENFORCE(impl != nullptr, "Allocator implementation cannot be null.");
+    impl_ = impl;
+  }
+
+  using AllocatorFactory = AllocatorPtr (*)(const OrtMemoryInfo& memory_info);
+
+  /**
+   * Create from an AllocatorFactory, which will be called lazily when the first allocation is made.
+   */
+  explicit Allocator(const OrtMemoryInfo* memory_info, AllocatorFactory get_allocator_impl)
+      : Allocator{memory_info} {
+    get_allocator_impl_ = get_allocator_impl;
+  }
+
+ private:
+  explicit Allocator(const OrtMemoryInfo* memory_info)
+      : OrtAllocator{}, memory_info_(memory_info) {
     version = ORT_API_VERSION;
     Alloc = AllocImpl;
     Free = FreeImpl;
     Info = InfoImpl;
   }
+  AllocatorPtr GetImpl() {
+    if (!impl_) {
+      std::call_once(init_flag_, [this]() {
+        impl_ = get_allocator_impl_(*memory_info_);
+      });
+    }
+    return impl_;
+  }
 
- private:
   static void* ORT_API_CALL AllocImpl(OrtAllocator* this_ptr, size_t size) noexcept {
     auto* allocator = static_cast<Allocator*>(this_ptr);
-    return allocator->impl_->Alloc(size);
+    return allocator->GetImpl()->Alloc(size);
   }
 
   static void ORT_API_CALL FreeImpl(OrtAllocator* this_ptr, void* p) noexcept {
     auto* allocator = static_cast<Allocator*>(this_ptr);
-    allocator->impl_->Free(p);
+    allocator->GetImpl()->Free(p);
   }
 
   static const OrtMemoryInfo* ORT_API_CALL InfoImpl(const OrtAllocator* this_ptr) noexcept {
@@ -44,6 +71,8 @@ class Allocator : public OrtAllocator {
 
   const OrtMemoryInfo* memory_info_;
   AllocatorPtr impl_;
+  AllocatorFactory get_allocator_impl_;
+  std::once_flag init_flag_;
 };
 
 }  // namespace adapter

include/onnxruntime/ep/adapter/ep.h

@@ -27,6 +27,7 @@ class Ep : public OrtEp {
         profiler_{impl_->GetProfiler()},
         temp_space_cpu_allocator_{temp_space_cpu_allocator},
         temp_space_allocator_{temp_space_allocator} {
+    ort_version_supported = ORT_API_VERSION;
   }
 
  public:

include/onnxruntime/ep/adapter/op_kernel_info.h

@@ -19,7 +19,7 @@
 #include "tensor_helper.h"
 
 namespace onnxruntime {
-struct DataTransferManager;
+class DataTransferManager;
 struct IExecutionProvider;
 }  // namespace onnxruntime
 

include/onnxruntime/ep/common.h

@@ -41,3 +41,24 @@
     OrtStatus* _status = (status_expr); \
     Ort::Status _ignored{_status};      \
   } while (false)
+
+// Helper macros to convert exceptions to OrtStatus* return values.
+// Usage:
+//   EXCEPTION_TO_RETURNED_STATUS_BEGIN
+//     ... code that may throw ...
+//   EXCEPTION_TO_RETURNED_STATUS_END
+#define EXCEPTION_TO_RETURNED_STATUS_BEGIN try {
+#define EXCEPTION_TO_RETURNED_STATUS_END                  \
+  }                                                       \
+  catch (const Ort::Exception& ex) {                      \
+    Ort::Status status(ex);                               \
+    return status.release();                              \
+  }                                                       \
+  catch (const std::exception& ex) {                      \
+    Ort::Status status(ex.what(), ORT_EP_FAIL);           \
+    return status.release();                              \
+  }                                                       \
+  catch (...) {                                           \
+    Ort::Status status("Unknown exception", ORT_EP_FAIL); \
+    return status.release();                              \
+  }