Adding DML to python cuda package (#22606)

onnxruntime/test/python/onnx_backend_test_series.py

@@ -105,7 +105,7 @@ def load_jsonc(basename: str):
     return json.loads("\n".join(lines))
 
 
-def create_backend_test(test_name=None):
+def create_backend_test(devices: list[str], test_name=None):
     """Creates an OrtBackendTest and adds its TestCase's to global scope so unittest will find them."""
 
     overrides = load_jsonc("onnx_backend_test_series_overrides.jsonc")
@@ -126,30 +126,29 @@ def create_backend_test(test_name=None):
     else:
         filters = load_jsonc("onnx_backend_test_series_filters.jsonc")
         current_failing_tests = apply_filters(filters, "current_failing_tests")
-
         if platform.architecture()[0] == "32bit":
             current_failing_tests += apply_filters(filters, "current_failing_tests_x86")
 
-        if backend.supports_device("DNNL"):
+        if backend.supports_device("DNNL") or "DNNL" in devices:
             current_failing_tests += apply_filters(filters, "current_failing_tests_DNNL")
 
-        if backend.supports_device("NNAPI"):
+        if backend.supports_device("NNAPI") or "NNAPI" in devices:
             current_failing_tests += apply_filters(filters, "current_failing_tests_NNAPI")
 
-        if backend.supports_device("OPENVINO_GPU"):
+        if backend.supports_device("OPENVINO_GPU") or "OPENVINO_GPU" in devices:
             current_failing_tests += apply_filters(filters, "current_failing_tests_OPENVINO_GPU")
 
-        if backend.supports_device("OPENVINO_CPU"):
+        if backend.supports_device("OPENVINO_CPU") or "OPENVINO_CPU" in devices:
             current_failing_tests += apply_filters(filters, "current_failing_tests_OPENVINO_CPU_FP32")
             current_failing_tests += apply_filters(filters, "current_failing_tests_OPENVINO_CPU_FP16")
 
-        if backend.supports_device("OPENVINO_NPU"):
+        if backend.supports_device("OPENVINO_NPU") or "OPENVINO_NPU" in devices:
             current_failing_tests += apply_filters(filters, "current_failing_tests_OPENVINO_NPU")
 
-        if backend.supports_device("OPENVINO"):
+        if backend.supports_device("OPENVINO") or "OPENVINO" in devices:
             current_failing_tests += apply_filters(filters, "current_failing_tests_OPENVINO_opset18")
 
-        if backend.supports_device("MIGRAPHX"):
+        if backend.supports_device("MIGRAPHX") or "MIGRAPHX" in devices:
             current_failing_tests += apply_filters(filters, "current_failing_tests_MIGRAPHX")
 
         if backend.supports_device("WEBGPU"):
@@ -158,8 +157,16 @@ def create_backend_test(test_name=None):
         # Skip these tests for a "pure" DML onnxruntime python wheel. We keep these tests enabled for instances where both DML and CUDA
         # EPs are available (Windows GPU CI pipeline has this config) - these test will pass because CUDA has higher precedence than DML
         # and the nodes are assigned to only the CUDA EP (which supports these tests)
-        if backend.supports_device("DML") and not backend.supports_device("GPU"):
+        if (backend.supports_device("DML") and not backend.supports_device("GPU")) or "DML" in devices:
             current_failing_tests += apply_filters(filters, "current_failing_tests_pure_DML")
+            # exclude CUDA EP when DML test is running.
+            os.environ["ORT_ONNX_BACKEND_EXCLUDE_PROVIDERS"] = "TensorrtExecutionProvider,CUDAExecutionProvider"
+        elif backend.supports_device("DML") and "DML" not in devices:
+            # exclude DML EP when CUDA test is running.
+            os.environ["ORT_ONNX_BACKEND_EXCLUDE_PROVIDERS"] = "TensorrtExecutionProvider,DmlExecutionProvider"
+        else:
+            # exclude TRT EP temporarily and only test CUDA EP to retain previous behavior
+            os.environ["ORT_ONNX_BACKEND_EXCLUDE_PROVIDERS"] = "TensorrtExecutionProvider"
 
         filters = (
             current_failing_tests
@@ -172,9 +179,6 @@ def create_backend_test(test_name=None):
         backend_test.exclude("(" + "|".join(filters) + ")")
         print("excluded tests:", filters)
 
-        # exclude TRT EP temporarily and only test CUDA EP to retain previous behavior
-        os.environ["ORT_ONNX_BACKEND_EXCLUDE_PROVIDERS"] = "TensorrtExecutionProvider"
-
     # import all test cases at global scope to make
     # them visible to python.unittest.
     globals().update(backend_test.enable_report().test_cases)
@@ -199,6 +203,15 @@ def parse_args():
         help="Only run tests that match this value. Matching is regex based, and '.*' is automatically appended",
     )
 
+    parser.add_argument(
+        "--devices",
+        type=str,
+        choices=["CPU", "CUDA", "MIGRAPHX", "DNNL", "DML", "OPENVINO_GPU", "OPENVINO_CPU", "OPENVINO_NPU", "OPENVINO"],
+        nargs="+",  # allows multiple values
+        default=["CPU"],  # default to ["CPU"] if no input is given
+        help="Select one or more devices CPU, CUDA, MIGRAPHX, DNNL, DML, OPENVINO_GPU, OPENVINO_CPU, OPENVINO_NPU, OPENVINO",
+    )
+
     # parse just our args. python unittest has its own args and arg parsing, and that runs inside unittest.main()
     parsed, unknown = parser.parse_known_args()
     sys.argv = sys.argv[:1] + unknown
@@ -209,5 +222,5 @@ def parse_args():
 if __name__ == "__main__":
     args = parse_args()
 
-    create_backend_test(args.test_name)
+    create_backend_test(args.devices, args.test_name)
     unittest.main()

onnxruntime/test/testdata/onnx_backend_test_series_filters.jsonc

@@ -750,6 +750,13 @@
         "^test_reduce_log_sum_empty_set_cpu",
         "^test_reduce_log_sum_exp_empty_set_cpu",
         "^test_reduce_prod_empty_set_cpu",
+        // Bug: DML EP some how executes these CUDA tests and failed
+        // TODO: Remove these tests when DML EP is fixed
+        "^test_convtranspose_autopad_same_cuda",
+        "^test_asin_example_cuda",
+        "^test_dynamicquantizelinear_cuda",
+        "^test_dynamicquantizelinear_expanded_cuda",
+        "^test_reduce_min_empty_set_cuda",
         //Bug: DML EP does not execute operators with an empty input tensor
         //TODO: Resolve as a graph implementation that returns a constant inf tensor with appropriate strides
         "^test_reduce_min_empty_set_cpu"

tools/ci_build/github/azure-pipelines/stages/jobs/steps/py_packaging_test_step.yml

@@ -0,0 +1,21 @@
+parameters:
+- name: EP_NAME
+  type: string
+  default: CPU
+
+- name: PYTHON_VERSION
+  type: string
+
+steps:
+- powershell: |
+    python -m pip uninstall -y onnxruntime onnxruntime-gpu -qq
+    Get-ChildItem -Path $(Build.ArtifactStagingDirectory)/*cp${{ replace(parameters.PYTHON_VERSION,'.','') }}*.whl | foreach {pip --disable-pip-version-check install --upgrade $_.fullname tabulate}
+    mkdir -p $(Agent.TempDirectory)\ort_test_data
+    Copy-Item -Path $(Build.sourcesDirectory)/onnxruntime/test/python/onnx_backend_test_series.py -Destination $(Agent.TempDirectory)\ort_test_data
+    Copy-Item -Recurse -Path $(Build.sourcesDirectory)/onnxruntime/test/testdata -Destination $(Agent.TempDirectory)\ort_test_data
+    cd $(Agent.TempDirectory)\ort_test_data
+    python onnx_backend_test_series.py --devices ${{ parameters.EP_NAME }} -v
+    cd $(Agent.TempDirectory)
+    Remove-Item -Path $(Agent.TempDirectory)\ort_test_data -Recurse -Force
+  workingDirectory: '$(Build.sourcesDirectory)'
+  displayName: 'Run Python Tests with ${{ parameters.EP_NAME }} EP'

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

@@ -56,7 +56,7 @@ stages:
           PYTHON_VERSION: ${{ python_version }}
           EP_NAME: gpu
           CudaVersion: ${{ parameters.cuda_version }}
-          EP_BUILD_FLAGS: --enable_lto --cuda_home=$(Agent.TempDirectory)\v${{ parameters.cuda_version }} --cmake_extra_defines "CMAKE_CUDA_ARCHITECTURES=52;60;61;70;75;80"
+          EP_BUILD_FLAGS: --use_dml --enable_lto --cuda_home=$(Agent.TempDirectory)\v${{ parameters.cuda_version }} --cmake_extra_defines "CMAKE_CUDA_ARCHITECTURES=52;60;61;70;75;80"
           use_tensorrt: True
 
   - ${{ if eq(parameters.enable_linux_cuda, true) }}:

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

@@ -33,7 +33,7 @@ parameters:
    - Release
    - RelWithDebInfo
    - MinSizeRel
-   
+
 - name: use_tensorrt
   type: boolean
   default: false
@@ -134,7 +134,7 @@ stages:
                 --cmake_generator "$(VSGenerator)"
                 --enable_pybind
                 --enable_onnx_tests
-                --parallel --use_binskim_compliant_compile_flags --update --build
+                --parallel 4 --use_binskim_compliant_compile_flags --update --build
                 $(TelemetryOption) ${{ parameters.BUILD_PY_PARAMETERS }} ${{ parameters.EP_BUILD_FLAGS }} ${{ variables.trt_build_flag }}
               workingDirectory: '$(Build.BinariesDirectory)'
 
@@ -206,19 +206,20 @@ stages:
             DownloadTRT: ${{ parameters.use_tensorrt }}
 
         - task: PowerShell@2
-          displayName: 'Install ONNX'
+          displayName: 'Install Third Party Dependencies'
           inputs:
             filePath: '$(Build.SourcesDirectory)/tools/ci_build/github/windows/install_third_party_deps.ps1'
             workingDirectory: '$(Build.BinariesDirectory)'
             arguments: -cpu_arch x64 -install_prefix $(Build.BinariesDirectory)\${{ parameters.cmake_build_type }}\installed -build_config ${{ parameters.cmake_build_type }}
 
-        - powershell: |
-            python -m pip uninstall -y onnxruntime onnxruntime-gpu -qq
-            Get-ChildItem -Path $(Build.ArtifactStagingDirectory)/*cp${{ replace(parameters.PYTHON_VERSION,'.','') }}*.whl | foreach {pip --disable-pip-version-check install --upgrade $_.fullname tabulate}
-            mkdir -p $(Agent.TempDirectory)\ort_test_data
-            Copy-Item -Path $(Build.sourcesDirectory)/onnxruntime/test/python/onnx_backend_test_series.py -Destination $(Agent.TempDirectory)\ort_test_data
-            Copy-Item -Recurse -Path $(Build.sourcesDirectory)/onnxruntime/test/testdata -Destination $(Agent.TempDirectory)\ort_test_data
-            cd $(Agent.TempDirectory)\ort_test_data
-            python onnx_backend_test_series.py
-          workingDirectory: '$(Build.sourcesDirectory)'
-          displayName: 'Run Python Tests'
+        - template: jobs/steps/py_packaging_test_step.yml
+          parameters:
+            EP_NAME: DML
+            PYTHON_VERSION: ${{ parameters.PYTHON_VERSION }}
+
+        - template: jobs/steps/py_packaging_test_step.yml
+          parameters:
+            EP_NAME: CUDA
+            PYTHON_VERSION: ${{ parameters.PYTHON_VERSION }}
+
+