Adding DML to python cuda package

include/onnxruntime/core/framework/ortdevice.h

@@ -17,6 +17,7 @@ struct OrtDevice {
   static const DeviceType GPU = 1;  // Nvidia or AMD
   static const DeviceType FPGA = 2;
   static const DeviceType NPU = 3;  // Ascend
+  static const DeviceType DML = 4;
 
   struct MemType {
     // Pre-defined memory types.

onnxruntime/core/providers/dml/DmlExecutionProvider/src/BucketizedBufferAllocator.cpp

@@ -47,7 +47,7 @@ namespace Dml
             OrtMemoryInfo(
                 "DML",
                 OrtAllocatorType::OrtDeviceAllocator,
-                OrtDevice(OrtDevice::GPU, OrtDevice::MemType::DEFAULT, 0)
+                OrtDevice(OrtDevice::DML, OrtDevice::MemType::DEFAULT, 0)
             )
         ),
         m_device(device),

onnxruntime/core/providers/dml/DmlExecutionProvider/src/DmlExternalBufferAllocator.h

@@ -24,7 +24,7 @@ namespace Dml
             OrtMemoryInfo(
                 "DML",
                 OrtAllocatorType::OrtDeviceAllocator,
-                OrtDevice(OrtDevice::GPU, OrtDevice::MemType::DEFAULT, 0)
+                OrtDevice(OrtDevice::DML, OrtDevice::MemType::DEFAULT, 0)
             ))
         {
             m_device = onnxruntime::DMLProviderFactoryCreator::CreateD3D12Device(device_id, false);

onnxruntime/core/providers/dml/DmlExecutionProvider/src/ExecutionProvider.cpp

@@ -74,7 +74,7 @@ namespace Dml
         bool enableGraphCapture,
         bool enableSyncSpinning,
         bool disableMemoryArena) :
-            IExecutionProvider(onnxruntime::kDmlExecutionProvider, OrtDevice(OrtDevice::GPU, OrtDevice::MemType::DEFAULT, 0))
+            IExecutionProvider(onnxruntime::kDmlExecutionProvider, OrtDevice(OrtDevice::DML, OrtDevice::MemType::DEFAULT, 0))
     {
         D3D12_COMMAND_LIST_TYPE queueType = executionContext->GetCommandListTypeForQueue();
         if (queueType != D3D12_COMMAND_LIST_TYPE_DIRECT && queueType != D3D12_COMMAND_LIST_TYPE_COMPUTE)

onnxruntime/core/providers/dml/DmlExecutionProvider/src/ExecutionProvider.h

@@ -242,8 +242,8 @@ namespace Dml
 
         bool CanCopy(const OrtDevice& srcDevice, const OrtDevice& dstDevice) const final
         {
-              return (srcDevice.Type() == OrtDevice::GPU) ||
-                     (dstDevice.Type() == OrtDevice::GPU);
+              return (srcDevice.Type() == OrtDevice::DML) ||
+                     (dstDevice.Type() == OrtDevice::DML);
         }
 
     private:

onnxruntime/python/onnxruntime_pybind_state.cc

@@ -288,11 +288,9 @@ const char* GetDeviceName(const OrtDevice& device) {
     case OrtDevice::CPU:
       return CPU;
     case OrtDevice::GPU:
-#ifdef USE_DML
-      return DML;
-#else
       return CUDA;
-#endif
+    case OrtDevice::DML:
+      return DML;
     case OrtDevice::FPGA:
       return "FPGA";
     case OrtDevice::NPU:
@@ -1579,7 +1577,7 @@ void addObjectMethods(py::module& m, ExecutionProviderRegistrationFn ep_registra
       .def_static("cann", []() { return OrtDevice::NPU; })
       .def_static("fpga", []() { return OrtDevice::FPGA; })
       .def_static("npu", []() { return OrtDevice::NPU; })
-      .def_static("dml", []() { return OrtDevice::GPU; })
+      .def_static("dml", []() { return OrtDevice::DML; })
       .def_static("default_memory", []() { return OrtDevice::MemType::DEFAULT; });
 
   py::class_<OrtArenaCfg> ort_arena_cfg_binding(m, "OrtArenaCfg");

tools/ci_build/github/azure-pipelines/stages/py-gpu-packaging-stage.yml

@@ -57,9 +57,9 @@ stages:
           EP_NAME: gpu
           CudaVersion: ${{ parameters.cuda_version }}
           ${{ if eq(parameters.cuda_version, '11.8') }}:
-            EP_BUILD_FLAGS: --enable_lto --use_tensorrt --tensorrt_home=$(Agent.TempDirectory)\TensorRT-10.4.0.26.Windows10.x86_64.cuda-11.8 --cuda_home=$(Agent.TempDirectory)\v11.8  --cmake_extra_defines "CMAKE_CUDA_ARCHITECTURES=52;60;61;70;75;80"
+            EP_BUILD_FLAGS: --use_dml --enable_wcos --enable_lto --use_tensorrt --tensorrt_home=$(Agent.TempDirectory)\TensorRT-10.4.0.26.Windows10.x86_64.cuda-11.8 --cuda_home=$(Agent.TempDirectory)\v11.8  --cmake_extra_defines "CMAKE_CUDA_ARCHITECTURES=52;60;61;70;75;80" --cmake_extra_defines "CMAKE_SYSTEM_VERSION=10.0.18362.0"
           ${{ if eq(parameters.cuda_version, '12.2') }}:
-            EP_BUILD_FLAGS: --enable_lto --use_tensorrt --tensorrt_home=$(Agent.TempDirectory)\TensorRT-10.4.0.26.Windows10.x86_64.cuda-12.6 --cuda_home=$(Agent.TempDirectory)\v12.2  --cmake_extra_defines "CMAKE_CUDA_ARCHITECTURES=52;60;61;70;75;80"
+            EP_BUILD_FLAGS: --use_dml --enable_wcos --enable_lto --use_tensorrt --tensorrt_home=$(Agent.TempDirectory)\TensorRT-10.4.0.26.Windows10.x86_64.cuda-12.6 --cuda_home=$(Agent.TempDirectory)\v12.2  --cmake_extra_defines "CMAKE_CUDA_ARCHITECTURES=52;60;61;70;75;80" --cmake_extra_defines "CMAKE_SYSTEM_VERSION=10.0.18362.0"
 
   - ${{ if eq(parameters.enable_linux_cuda, true) }}:
       - template: py-linux-gpu-stage.yml