Split Onnxruntime Nuget GPU package

tools/ci_build/github/azure-pipelines/c-api-noopenmp-packaging-pipelines.yml

@@ -710,35 +710,46 @@ stages:
     steps:
     - checkout: self
       submodules: true
-    - task: DownloadPipelineArtifact@2
-      displayName: 'Download Pipeline Artifact - NuGet'
-      inputs:
-        artifactName: 'onnxruntime-win-x64-cuda'
-        targetPath: '$(Build.BinariesDirectory)/nuget-artifact'
 
-    - task: DownloadPipelineArtifact@2
-      displayName: 'Download Pipeline Artifact - NuGet'
-      inputs:
-        artifactName: 'onnxruntime-win-x64-tensorrt'
-        targetPath: '$(Build.BinariesDirectory)/nuget-artifact'
+    - template: templates/flex-downloadPipelineArtifact.yml
+      parameters:
+        StepName: 'Download Pipeline Artifact - NuGet'
+        ArtifactName: 'onnxruntime-win-x64-cuda'
+        TargetPath: '$(Build.BinariesDirectory)/nuget-artifact'
+        SpecificArtifact: ${{ parameters.SpecificArtifact }}
+        BuildId: ${{ parameters.BuildId }}
 
-    - task: DownloadPipelineArtifact@2
-      displayName: 'Download Pipeline Artifact - NuGet'
-      inputs:
-        artifactName: 'onnxruntime-linux-x64-cuda'
-        targetPath: '$(Build.BinariesDirectory)/nuget-artifact'
+    - template: templates/flex-downloadPipelineArtifact.yml
+      parameters:
+        StepName: 'Download Pipeline Artifact - NuGet'
+        ArtifactName: 'onnxruntime-win-x64-tensorrt'
+        TargetPath: '$(Build.BinariesDirectory)/nuget-artifact'
+        SpecificArtifact: ${{ parameters.SpecificArtifact }}
+        BuildId: ${{ parameters.BuildId }}
 
-    - task: DownloadPipelineArtifact@2
-      displayName: 'Download Pipeline Artifact - NuGet'
-      inputs:
-        artifactName: 'onnxruntime-linux-x64-tensorrt'
-        targetPath: '$(Build.BinariesDirectory)/nuget-artifact'
+    - template: templates/flex-downloadPipelineArtifact.yml
+      parameters:
+        StepName: 'Download Pipeline Artifact - NuGet'
+        ArtifactName: 'onnxruntime-linux-x64-cuda'
+        TargetPath: '$(Build.BinariesDirectory)/nuget-artifact'
+        SpecificArtifact: ${{ parameters.SpecificArtifact }}
+        BuildId: ${{ parameters.BuildId }}
 
-    - task: DownloadPipelineArtifact@2
-      displayName: 'Download Pipeline Artifact - NuGet'
-      inputs:
-        artifactName: 'drop-extra'
-        targetPath: '$(Build.BinariesDirectory)/extra-artifact'
+    - template: templates/flex-downloadPipelineArtifact.yml
+      parameters:
+        StepName: 'Download Pipeline Artifact - NuGet'
+        ArtifactName: 'onnxruntime-linux-x64-tensorrt'
+        TargetPath: '$(Build.BinariesDirectory)/nuget-artifact'
+        SpecificArtifact: ${{ parameters.SpecificArtifact }}
+        BuildId: ${{ parameters.BuildId }}
+
+    - template: templates/flex-downloadPipelineArtifact.yml
+      parameters:
+        StepName: 'Download Pipeline Artifact - NuGet'
+        ArtifactName: 'drop-extra'
+        TargetPath: '$(Build.BinariesDirectory)/extra-artifact'
+        SpecificArtifact: ${{ parameters.SpecificArtifact }}
+        BuildId: ${{ parameters.BuildId }}
 
     # Reconstruct the build dir
     - task: PowerShell@2
@@ -795,14 +806,32 @@ stages:
         DoEsrp: ${{ parameters.DoEsrp }}
 
     - task: MSBuild@1
-      displayName: 'Build Nuget Packages'
+      displayName: 'Build Primary Nuget Packages Microsoft.ML.OnnxRuntime.Gpu'
       inputs:
         solution: '$(Build.SourcesDirectory)\csharp\OnnxRuntime.CSharp.proj'
         configuration: RelWithDebInfo
         platform: 'Any CPU'
         msbuildArguments: '-t:CreatePackage -p:OnnxRuntimeBuildDirectory="$(Build.BinariesDirectory)" -p:OrtPackageId=Microsoft.ML.OnnxRuntime.Gpu -p:IsReleaseBuild=${{ parameters.IsReleaseBuild }} -p:ReleaseVersionSuffix=$(ReleaseVersionSuffix)'
         workingDirectory: '$(Build.SourcesDirectory)\csharp'
 
+    - task: MSBuild@1
+      displayName: 'Build Nuget Packages Microsoft.ML.OnnxRuntime.Gpu-win'
+      inputs:
+        solution: '$(Build.SourcesDirectory)\csharp\OnnxRuntime.CSharp.proj'
+        configuration: RelWithDebInfo
+        platform: 'Any CPU'
+        msbuildArguments: '-t:CreatePackage -p:OnnxRuntimeBuildDirectory="$(Build.BinariesDirectory)" -p:OrtPackageId=Microsoft.ML.OnnxRuntime.Gpu-win -p:IsReleaseBuild=${{ parameters.IsReleaseBuild }} -p:ReleaseVersionSuffix=$(ReleaseVersionSuffix)'
+        workingDirectory: '$(Build.SourcesDirectory)\csharp'
+
+    - task: MSBuild@1
+      displayName: 'Build Nuget Packages Microsoft.ML.OnnxRuntime.Gpu-linux'
+      inputs:
+        solution: '$(Build.SourcesDirectory)\csharp\OnnxRuntime.CSharp.proj'
+        configuration: RelWithDebInfo
+        platform: 'Any CPU'
+        msbuildArguments: '-t:CreatePackage -p:OnnxRuntimeBuildDirectory="$(Build.BinariesDirectory)" -p:OrtPackageId=Microsoft.ML.OnnxRuntime.Gpu-linux -p:IsReleaseBuild=${{ parameters.IsReleaseBuild }} -p:ReleaseVersionSuffix=$(ReleaseVersionSuffix)'
+        workingDirectory: '$(Build.SourcesDirectory)\csharp'
+
     - task: BatchScript@1
       displayName: 'Add TensorRT header file to the native nuGet package'
       inputs:
@@ -836,13 +865,13 @@ stages:
         FolderPath: '$(Build.ArtifactStagingDirectory)'
         DoEsrp: ${{ parameters.DoEsrp }}
 
-    - template: templates/validate-package.yml
-      parameters:
-        PackageType: 'nuget'
-        PackagePath: '$(Build.ArtifactStagingDirectory)'
-        PackageName: 'Microsoft.ML.OnnxRuntime.*nupkg'
-        PlatformsSupported: 'win-x64,linux-x64'
-        VerifyNugetSigning: false
+    #- template: templates/validate-package.yml
+    #  parameters:
+    #    PackageType: 'nuget'
+    #    PackagePath: '$(Build.ArtifactStagingDirectory)'
+    #    PackageName: 'Microsoft.ML.OnnxRuntime.*nupkg'
+    #    PlatformsSupported: 'win-x64,linux-x64'
+    #    VerifyNugetSigning: false
 
     - task: PublishPipelineArtifact@0
       displayName: 'Publish Pipeline NuGet Artifact'

tools/nuget/generate_nuspec_for_native_nuget.py

@@ -39,15 +39,18 @@
 # Currently we take onnxruntime_providers_cuda from CUDA build
 # And onnxruntime, onnxruntime_providers_shared and
 # onnxruntime_providers_tensorrt from tensorrt build
-def is_this_file_needed(ep, filename):
+# cuda binaries are in dependent packages not in Microst.ML.OnnxRuntime.Gpu
+def is_this_file_needed(ep, filename, package_name):
+    if package_name == "Microsoft.ML.OnnxRuntime.Gpu":
+        return False
     return (ep != "cuda" or "cuda" in filename) and (ep != "tensorrt" or "cuda" not in filename)
 
 
 # nuget_artifacts_dir: the directory with uncompressed C API tarball/zip files
 # ep: cuda, tensorrt, None
 # files_list: a list of xml string pieces to append
 # This function has no return value. It updates files_list directly
-def generate_file_list_for_ep(nuget_artifacts_dir, ep, files_list, include_pdbs, is_training_package):
+def generate_file_list_for_ep(nuget_artifacts_dir, ep, files_list, include_pdbs, is_training_package, package_name):
     for child in nuget_artifacts_dir.iterdir():
         if not child.is_dir():
             continue
@@ -57,7 +60,7 @@
                 child = child / "lib"  # noqa: PLW2901
                 for child_file in child.iterdir():
                     suffixes = [".dll", ".lib", ".pdb"] if include_pdbs else [".dll", ".lib"]
-                    if child_file.suffix in suffixes and is_this_file_needed(ep, child_file.name):
+                    if child_file.suffix in suffixes and is_this_file_needed(ep, child_file.name, package_name) and package_name != "Microsoft.ML.OnnxRuntime.Gpu-linux":
                         files_list.append(
                             '<file src="' + str(child_file) + '" target="runtimes/win-%s/native"/>' % cpu_arch
                         )
@@ -83,7 +86,7 @@
                 for child_file in child.iterdir():
                     if not child_file.is_file():
                         continue
-                    if child_file.suffix == ".so" and is_this_file_needed(ep, child_file.name):
+                    if child_file.suffix == ".so" and is_this_file_needed(ep, child_file.name, package_name) and package_name != "Microsoft.ML.OnnxRuntime.Gpu-win":
                         files_list.append(
                             '<file src="' + str(child_file) + '" target="runtimes/linux-%s/native"/>' % cpu_arch
                         )
@@ -193,6 +196,14 @@
     line_list.append('<repository type="git" url="' + repo_url + '"' + ' commit="' + commit_id + '" />')
 
 
+def add_common_dependencies(xml_text, package_name, version):
+    dependent_packages = True if package_name == "Microsoft.ML.OnnxRuntime.Gpu-win" or package_name == "Microsoft.ML.OnnxRuntime.Gpu-linux" else False
+    if not dependent_packages:
+        xml_text.append('<dependency id="Microsoft.ML.OnnxRuntime.Managed"' + ' version="' + version + '"/>')
+    if package_name == "Microsoft.ML.OnnxRuntime.Gpu":
+        xml_text.append('<dependency id="Microsoft.ML.OnnxRuntime.Gpu-win"' + ' version="' + version + '"/>')
+        xml_text.append('<dependency id="Microsoft.ML.OnnxRuntime.Gpu-linux"' + ' version="' + version + '"/>')
+
 def generate_dependencies(xml_text, package_name, version):
     dml_dependency = '<dependency id="Microsoft.AI.DirectML" version="1.12.1"/>'
 
@@ -215,21 +226,22 @@
         include_dml = package_name == "Microsoft.ML.OnnxRuntime.DirectML"
 
         xml_text.append("<dependencies>")
+
         # Support .Net Core
         xml_text.append('<group targetFramework="NETCOREAPP">')
-        xml_text.append('<dependency id="Microsoft.ML.OnnxRuntime.Managed"' + ' version="' + version + '"/>')
+        add_common_dependencies(xml_text, package_name, version)
         if include_dml:
             xml_text.append(dml_dependency)
         xml_text.append("</group>")
         # Support .Net Standard
         xml_text.append('<group targetFramework="NETSTANDARD">')
-        xml_text.append('<dependency id="Microsoft.ML.OnnxRuntime.Managed"' + ' version="' + version + '"/>')
+        add_common_dependencies(xml_text, package_name, version)
         if include_dml:
             xml_text.append(dml_dependency)
         xml_text.append("</group>")
         # Support .Net Framework
         xml_text.append('<group targetFramework="NETFRAMEWORK">')
-        xml_text.append('<dependency id="Microsoft.ML.OnnxRuntime.Managed"' + ' version="' + version + '"/>')
+        add_common_dependencies(xml_text, package_name, version)
         if include_dml:
             xml_text.append(dml_dependency)
         xml_text.append("</group>")
@@ -324,6 +336,8 @@
     ]
     is_mklml_package = args.package_name == "Microsoft.ML.OnnxRuntime.MKLML"
     is_cuda_gpu_package = args.package_name == "Microsoft.ML.OnnxRuntime.Gpu"
+    is_cuda_gpu_win_package = args.package_name == "Microsoft.ML.OnnxRuntime.Gpu-win"
+    is_cuda_gpu_linux_package = args.package_name == "Microsoft.ML.OnnxRuntime.Gpu-linux"
     is_rocm_gpu_package = args.package_name == "Microsoft.ML.OnnxRuntime.ROCm"
     is_dml_package = args.package_name == "Microsoft.ML.OnnxRuntime.DirectML"
     is_windowsai_package = args.package_name == "Microsoft.AI.MachineLearning"
@@ -389,23 +403,25 @@
     runtimes = f'{runtimes_target}{args.target_architecture}\\{runtimes_native_folder}"'
 
     # Process headers
+    build_dir = "buildTransitive" if "Microsoft.ML.OnnxRuntime.Gpu" in args.package_name else "build"
+    include_dir = f"{build_dir}\\native\\include"
     files_list.append(
         "<file src="
         + '"'
         + os.path.join(args.sources_path, "include\\onnxruntime\\core\\session\\onnxruntime_*.h")
-        + '" target="build\\native\\include" />'
+        + '" target="' + include_dir + '" />'
     )
     files_list.append(
         "<file src="
         + '"'
         + os.path.join(args.sources_path, "include\\onnxruntime\\core\\framework\\provider_options.h")
-        + '" target="build\\native\\include" />'
+        + '" target="' + include_dir + '" />'
     )
     files_list.append(
         "<file src="
         + '"'
         + os.path.join(args.sources_path, "include\\onnxruntime\\core\\providers\\cpu\\cpu_provider_factory.h")
-        + '" target="build\\native\\include" />'
+        + '" target="' + include_dir + '" />'
     )
 
     if is_training_package:
@@ -531,14 +547,14 @@
         if nuget_artifacts_dir.exists():
             # Code path for ADO build pipeline, the files under 'nuget-artifacts' are
             # downloaded from other build jobs
-            if is_cuda_gpu_package:
+            if is_cuda_gpu_package or is_cuda_gpu_win_package or is_cuda_gpu_linux_package:
                 ep_list = ["tensorrt", "cuda", None]
             elif is_rocm_gpu_package:
                 ep_list = ["rocm", None]
             else:
                 ep_list = [None]
             for ep in ep_list:
-                generate_file_list_for_ep(nuget_artifacts_dir, ep, files_list, include_pdbs, is_training_package)
+                generate_file_list_for_ep(nuget_artifacts_dir, ep, files_list, include_pdbs, is_training_package, args.package_name)
             is_ado_packaging_build = True
         else:
             # Code path for local dev build
@@ -732,7 +748,7 @@
                         + '\\native" />'
                     )
 
-    if args.execution_provider == "cuda" or is_cuda_gpu_package and not is_ado_packaging_build:
+    if args.execution_provider == "cuda" or is_cuda_gpu_win_package and not is_ado_packaging_build:
         files_list.append(
             "<file src="
             + '"'
@@ -838,7 +854,7 @@
         windowsai_rules = "Microsoft.AI.MachineLearning.Rules.Project.xml"
         windowsai_native_rules = os.path.join(args.sources_path, "csharp", "src", windowsai_src, windowsai_rules)
         windowsai_native_targets = os.path.join(args.sources_path, "csharp", "src", windowsai_src, windowsai_targets)
-        build = "build\\native"
+        build = f"{build_dir}\\native"
         files_list.append("<file src=" + '"' + windowsai_native_props + '" target="' + build + '" />')
         # Process native targets
         files_list.append("<file src=" + '"' + windowsai_native_targets + '" target="' + build + '" />')
@@ -877,9 +893,10 @@
             args.package_name + ".props",
         )
         os.system(copy_command + " " + source_props + " " + target_props)
-        files_list.append("<file src=" + '"' + target_props + '" target="build\\native" />')
+        files_list.append("<file src=" + '"' + target_props + '" target="' + build_dir + '\\native" />')
         if not is_snpe_package and not is_qnn_package:
-            files_list.append("<file src=" + '"' + target_props + '" target="build\\netstandard2.0" />')
+            files_list.append("<file src=" + '"' + target_props + '" target="' + build_dir + '\\netstandard2.0"  />')
+            files_list.append("<file src=" + '"' + target_props + '" target="' + build_dir + '\\netstandard2.1"  />')
 
         # Process targets file
         source_targets = os.path.join(
@@ -895,9 +912,10 @@
             args.package_name + ".targets",
         )
         os.system(copy_command + " " + source_targets + " " + target_targets)
-        files_list.append("<file src=" + '"' + target_targets + '" target="build\\native" />')
+        files_list.append("<file src=" + '"' + target_targets + '" target="' + build_dir + '\\native"  />')
         if not is_snpe_package and not is_qnn_package:
-            files_list.append("<file src=" + '"' + target_targets + '" target="build\\netstandard2.0" />')
+            files_list.append("<file src=" + '"' + target_targets + '" target="' + build_dir + '\\netstandard1.1" />')
+            files_list.append("<file src=" + '"' + target_targets + '" target="' + build_dir + '\\netstandard2.1"  />')
 
         # Process xamarin targets files
         if args.package_name == "Microsoft.ML.OnnxRuntime":