Build onnxruntime-genai-cuda packages on win-arm64

.pipelines/mac-cpu-arm64-build.yml

@@ -190,12 +190,6 @@ jobs:
     displayName: 'Run the Java tests'
     workingDirectory: '$(Build.SourcesDirectory)'
 
-  - bash: |
-      set -e -x
-      rm -rf ort
-    displayName: 'Remove the ort lib and header files'
-    workingDirectory: '$(Build.SourcesDirectory)'
-
   - bash: |
       ls -l $(Build.SourcesDirectory)/build/cpu/osx-arm64
       ls -l $(Build.SourcesDirectory)/build/cpu/osx-arm64/wheel
@@ -216,7 +210,7 @@ jobs:
       set -e -x
       export ORTGENAI_LOG_ORT_LIB=1
       cd test/csharp
-      dotnet test /p:Configuration=Release /p:NativeBuildOutputDir="../../build/cpu/osx-arm64" --verbosity normal
+      dotnet test /p:Configuration=Release /p:NativeBuildOutputDir="../../build/cpu/osx-arm64" /p:OrtLibDir="../../ort/lib" --verbosity normal
     displayName: 'Build the C# API and run the C# tests'
     workingDirectory: '$(Build.SourcesDirectory)'
 

.pipelines/nuget-publishing.yml

@@ -83,7 +83,7 @@ parameters:
     type: string
     values:
       - '12.8'
-      - '13.0'
+      - '13.1'
     default: '12.8'
 
   - name: build_config

.pipelines/pypl-publishing.yml

@@ -19,8 +19,8 @@ parameters:
     type: boolean
     default: true
 
-  - name: enable_win_arm64_cpu
-    displayName: 'Whether Windows ARM64 CPU package is built.'
+  - name: enable_win_arm64
+    displayName: 'Whether Windows ARM64 package is built.'
     type: boolean
     default: true
 
@@ -54,6 +54,11 @@ parameters:
     type: string
     default: '2.1.1'
 
+  - name: ort_cuda_version
+    displayName: 'OnnxRuntime GPU version'
+    type: string
+    default: '1.23.0'
+
   - name: cuda_versions
     displayName: 'CUDA versions'
     type: string
@@ -115,11 +120,12 @@ extends:
           enable_win_x64_cpu: ${{ parameters.enable_win_x64_cpu }}
           enable_win_cuda: ${{ parameters.enable_win_cuda }}
           enable_win_dml: ${{ parameters.enable_win_dml }}
-          enable_win_arm64_cpu: ${{ parameters.enable_win_arm64_cpu }}
+          enable_win_arm64: ${{ parameters.enable_win_arm64 }}
           enable_macos_cpu: ${{ parameters.enable_macos_cpu }}
           cuda_versions: ${{ parameters.cuda_versions }}
           build_config: ${{ parameters.build_config }}
           enable_winml: ${{ parameters.enable_winml }}
+          ort_cuda_version: ${{ parameters.ort_cuda_version }}
           ${{ if eq(parameters.enable_winml, true) }}:
             ort_version: ${{ parameters.ort_winml_version }}
           ${{ else }}:
@@ -136,7 +142,7 @@ extends:
               enable_win_x64_cpu: ${{ parameters.enable_win_x64_cpu }}
               enable_win_cuda: ${{ parameters.enable_win_cuda }}
               enable_win_dml: ${{ parameters.enable_win_dml }}
-              enable_win_arm64_cpu: ${{ parameters.enable_win_arm64_cpu }}
+              enable_win_arm64: ${{ parameters.enable_win_arm64 }}
               enable_macos_cpu: ${{ parameters.enable_macos_cpu }}
               cuda_versions: ${{ parameters.cuda_versions }}
               SpecificArtifact: false

.pipelines/stages/capi-packaging-stage.yml

@@ -67,6 +67,7 @@ stages:
         enable_winml: ${{ parameters.enable_winml }}
         ort_version: ${{ parameters.ort_winml_version }}
         build_config: ${{ parameters.build_config }}
+        cuda_version: ${{ parameters.cuda_version }}
 
   - ${{ if eq(parameters.enable_win_x64_cpu, true) }}:
     - template: jobs/capi-packaging-job.yml
@@ -129,6 +130,19 @@ stages:
         ort_build_version: ${{ parameters.ort_build_version }}
         ort_build_source: ${{ parameters.ort_build_source }}
 
+  - ${{ if and(eq(parameters.enable_win_arm64, true), eq(parameters.enable_win_cuda, true)) }}:
+    - template: jobs/capi-packaging-job.yml
+      parameters:
+        arch: 'arm64'
+        cuda_version: ${{ parameters.cuda_version }}
+        ep: 'cuda'
+        ort_version: ${{ parameters.ort_cuda_version }}
+        os: 'win'
+        build_config: ${{ parameters.build_config }}
+        enable_winml: ${{ parameters.enable_winml }}
+        ort_build_version: ${{ parameters.ort_build_version }}
+        ort_build_source: ${{ parameters.ort_build_source }}
+
   - ${{ if eq(parameters.enable_linux_x64_cpu, true) }}:
     - template: jobs/capi-packaging-job.yml
       parameters:

.pipelines/stages/jobs/custom-nuget-packaging-job.yml

@@ -22,8 +22,8 @@ jobs:
 
   - template: steps/utils/flex-download-pipeline-artifact.yml
     parameters:
-      StepName: 'Download win-cpu-arm64 capi Artifacts'
-      ArtifactName: onnxruntime-genai-win-cpu-arm64-capi
+      StepName: 'Download win-cuda-arm64 capi Artifacts'
+      ArtifactName: onnxruntime-genai-win-cuda-arm64-capi
       TargetPath: '$(Build.BinariesDirectory)/artifact-downloads'
 
   - template: steps/utils/flex-download-pipeline-artifact.yml

.pipelines/stages/jobs/nuget-packaging-job.yml

@@ -144,6 +144,13 @@ jobs:
         ArtifactName: onnxruntime-genai-win-cuda-x64-capi
         TargetPath: '$(Build.BinariesDirectory)/artifact-downloads'
 
+  - ${{ if and(eq(parameters.enable_win_arm64, true), eq(parameters.enable_win_cuda, true), eq(parameters.ep, 'cuda')) }}:
+    - template: steps/utils/flex-download-pipeline-artifact.yml
+      parameters:
+        StepName: 'Download win-cuda-arm64 capi Artifacts'
+        ArtifactName: onnxruntime-genai-win-cuda-arm64-capi
+        TargetPath: '$(Build.BinariesDirectory)/artifact-downloads'
+
   - ${{ if and(eq(parameters.enable_win_dml, true), eq(parameters.ep, 'directml')) }}:
     - template: steps/utils/flex-download-pipeline-artifact.yml
       parameters:
@@ -317,6 +324,19 @@ jobs:
         }
       }
 
+      if ('${{ parameters.ep }}' -eq 'cuda') {
+        $cudaExpected = @{}
+        if ('${{ parameters.enable_win_cuda }}' -eq 'True')   { $cudaExpected['win-x64']   = 'onnxruntime-genai-cuda.dll' }
+        if ('${{ parameters.enable_win_arm64 }}' -eq 'True')  { $cudaExpected['win-arm64'] = 'onnxruntime-genai-cuda.dll' }
+        if ('${{ parameters.enable_linux_cuda }}' -eq 'True') { $cudaExpected['linux-x64'] = 'libonnxruntime-genai-cuda.so' }
+        foreach ($rid in $cudaExpected.Keys) {
+          $lib = "$outputDir/$rid/$(buildConfig)/$($cudaExpected[$rid])"
+          if (-not (Test-Path $lib)) {
+            throw "Expected CUDA native library missing for enabled RID '$rid': $lib"
+          }
+        }
+      }
+
     displayName: 'Extract Artifacts & Prepare Native Libraries'
     workingDirectory: '$(Build.BinariesDirectory)/artifact-downloads'
 

.pipelines/stages/jobs/nuget-validation-job.yml

@@ -111,7 +111,7 @@ jobs:
       value: 'cpu_and_mobile/cpu-int4-rtn-block-32-acc-level-4'
 
   - name: cuda_docker_image
-    ${{ if eq(parameters.cuda_version, '13.0') }}:
+    ${{ if eq(parameters.cuda_version, '13.1') }}:
       value: onnxruntimebuildcache.azurecr.io/internal/azureml/onnxruntime/build/cuda13_x64_almalinux8_gcc14:20251107.1
     ${{ else }}:
       value: onnxruntimebuildcache.azurecr.io/internal/azureml/onnxruntime/build/cuda12_x64_almalinux8_gcc12:20250714.2

.pipelines/stages/jobs/py-validation-job.yml

@@ -124,7 +124,7 @@ jobs:
       value: 'cpu_and_mobile/cpu-int4-rtn-block-32-acc-level-4'
 
   - name: cuda_docker_image
-    ${{ if eq(parameters.cuda_version, '13.0') }}:
+    ${{ if eq(parameters.cuda_version, '13.1') }}:
       value: onnxruntimebuildcache.azurecr.io/internal/azureml/onnxruntime/build/cuda13_x64_almalinux8_gcc14:20251107.1
     ${{ else }}:
       value: onnxruntimebuildcache.azurecr.io/internal/azureml/onnxruntime/build/cuda12_x64_almalinux8_gcc12:20250714.2

.pipelines/stages/jobs/steps/capi-linux-step.yml

@@ -52,7 +52,11 @@ steps:
 - bash: |
     set -e -x
     python3 -m pip install requests
-    python3 tools/ci_build/get_docker_image.py --dockerfile tools/ci_build/github/linux/docker/manylinux/Dockerfile.manylinux2_28_$(ep)_$(cuda_version) \
+    cuda_dockerfile_version="$(cuda_version)"
+    if [[ "$(cuda_version)" == 13.* ]]; then
+      cuda_dockerfile_version="13"
+    fi
+    python3 tools/ci_build/get_docker_image.py --dockerfile tools/ci_build/github/linux/docker/manylinux/Dockerfile.manylinux2_28_$(ep)_$cuda_dockerfile_version \
       --context tools/ci_build/github/linux/docker/manylinux \
       --docker-build-args "--build-arg BUILD_UID=$( id -u )" \
       --container-registry onnxruntimebuildcache \
@@ -161,15 +165,28 @@ steps:
 - ${{ if eq(parameters.target, 'onnxruntime-genai') }}:
   - script: |
       set -e -x
+      docker_env_args=()
+      if [ "$(ep)" = "cuda" ]; then
+        docker_env_args=(-e CMAKE_EXTRA_ARGS="$(CMAKE_EXTRA_ARGS)")
+      fi
       docker run \
       --rm \
       --volume $(Build.Repository.LocalPath):/ort_genai_src \
-      -w /ort_genai_src/ ortgenai$(ep)build$(arch) \
-      bash -c " \
-          /usr/bin/cmake --preset linux_gcc_$(ep)_$(build_config) \
-            -DENABLE_TESTS=OFF -DENABLE_PYTHON=OFF $(CMAKE_EXTRA_ARGS) $(ORT_HOME_CMAKE_ARG) && \
+      -w /ort_genai_src/ \
+      "${docker_env_args[@]}" \
+      ortgenai$(ep)build$(arch) \
+      bash -c '
+          set -e -x
+          cmake_args=(-DENABLE_TESTS=OFF -DENABLE_PYTHON=OFF)
+          if [ -n "$CMAKE_EXTRA_ARGS" ]; then
+            eval "cmake_args+=($CMAKE_EXTRA_ARGS)"
+          fi
+          if [ "${{ parameters.ort_build_version }}" != "default" ]; then
+            cmake_args+=("-DORT_HOME=/ort_genai_src/ort_build")
+          fi
+          /usr/bin/cmake --preset linux_gcc_$(ep)_$(build_config) "${cmake_args[@]}"
           /usr/bin/cmake --build --preset linux_gcc_$(ep)_$(build_config) \
-            --target onnxruntime-genai"
+            --target onnxruntime-genai'
     displayName: 'Build GenAi'
     workingDirectory: '$(Build.Repository.LocalPath)'
 
@@ -252,4 +269,3 @@ steps:
 - script: |
     ls $(Build.Repository.LocalPath) -R
   displayName: 'List files from SourceDirectory'
-

.pipelines/stages/jobs/steps/capi-win-step.yml

@@ -64,10 +64,62 @@ steps:
 - powershell: |
     azcopy.exe cp --recursive "https://lotusscus.blob.core.windows.net/models/cuda_sdk/v$(cuda_version)" 'cuda_sdk'
   displayName: 'Download CUDA $(cuda_version)'
-  condition: or(eq(variables['ep'], 'cuda'), and(eq(variables['ep'], 'winml'), eq(variables['arch'], 'x64')))
+  condition: and(eq(variables['arch'], 'x64'), or(eq(variables['ep'], 'cuda'), eq(variables['ep'], 'winml')))
   workingDirectory: '$(Build.Repository.LocalPath)'
 
-- ${{ if or(eq(parameters.ep, 'cuda'), and(eq(parameters.ep, 'winml'), eq(parameters.arch, 'x64'))) }}:
+- powershell: |
+    Write-Host "Downloading Azure CLI..."
+    $azCliUrl = "https://aka.ms/installazurecliwindows"
+    $installerPath = "$(Agent.TempDirectory)\AzureCLI.msi"
+    Invoke-WebRequest -Uri $azCliUrl -OutFile $installerPath
+    Write-Host "Installing Azure CLI..."
+    Start-Process -FilePath "msiexec.exe" -ArgumentList "/i", $installerPath, "/quiet", "/norestart" -Wait
+    Write-Host "Azure CLI installed successfully"
+    # Add Azure CLI to PATH
+    $azCliBin = "C:\Program Files (x86)\Microsoft SDKs\Azure\CLI2\wbin"
+    if (!(Test-Path $azCliBin)) {
+      Write-Error "Azure CLI install failed or path not found: $azCliBin"
+      exit 1
+    }
+    $env:PATH = "$azCliBin;$env:PATH"
+
+    Write-Host "Downloading AzCopy..."
+    $azcopyUrl = "https://aka.ms/downloadazcopy-v10-windows"
+    $azcopyZip = "$(Agent.TempDirectory)\azcopy.zip"
+    Invoke-WebRequest -Uri $azcopyUrl -OutFile $azcopyZip
+    $azcopyExtractPath = "$(Agent.TempDirectory)\azcopy"
+    Expand-Archive -Path $azcopyZip -DestinationPath $azcopyExtractPath -Force
+    # Find the actual azcopy.exe (it's in a versioned subfolder)
+    $azcopyExeFile = Get-ChildItem -Path $azcopyExtractPath -Name "azcopy.exe" -Recurse | Select-Object -First 1
+    if (-not $azcopyExeFile) {
+      Write-Error "AzCopy executable not found after extraction"
+      exit 1
+    }
+    $azcopyFullPath = Get-ChildItem -Path $azcopyExtractPath -Filter "azcopy.exe" -Recurse | Select-Object -First 1
+    $azcopyDir = $azcopyFullPath.Directory.FullName
+    Write-Host "AzCopy found at: $azcopyDir"
+    $env:PATH = "$azcopyDir;$env:PATH"
+
+    Write-Host "##vso[task.setvariable variable=PATH]$env:PATH"
+    # Verify installations
+    az --version
+    azcopy --version
+  displayName: 'Install Azure CLI and AzCopy'
+  condition: and(eq(variables['arch'], 'arm64'), or(eq(variables['ep'], 'winml'), eq(variables['ep'], 'cuda')))
+  workingDirectory: '$(Build.Repository.LocalPath)'
+
+- ${{ if and(or(eq(parameters.ep, 'cuda'), eq(parameters.ep, 'winml')), eq(parameters.arch, 'arm64')) }}:
+  - task: AzureCLI@2
+    displayName: 'Download CUDA $(cuda_version)'
+    inputs:
+      azureSubscription: 'ortcibuild_readonly_mi2-ONNX Runtime'
+      scriptType: batch
+      scriptLocation: inlineScript
+      inlineScript: |
+        set AZCOPY_AUTO_LOGIN_TYPE=AZCLI
+        azcopy cp --recursive "https://lotusscus.blob.core.windows.net/models/cuda_sdk/win-arm64/v$(cuda_version)" "cuda_sdk"
+
+- ${{ if or(eq(parameters.ep, 'cuda'), eq(parameters.ep, 'winml')) }}:
   - powershell: |
       if ('$(cuda_version)' -like '13.*') {
         $cuda_archs = "75-real;80-real;86-real;89-real;90-real;100-real;120-real;120-virtual"
@@ -108,13 +160,13 @@ steps:
   condition: eq(variables['arch'], 'arm64')
   workingDirectory: '$(Build.Repository.LocalPath)'
 
-- ${{ if ne(parameters.ort_build_version, 'default') }}:
+- ${{ if or(ne(parameters.ort_build_version, 'default'), and(eq(parameters.ep, 'cuda'), eq(parameters.arch, 'arm64'))) }}:
   - powershell: |
       $ep = "$(ep)"
       $arch = "$(arch)"
 
       switch ($ep) {
-          "cuda" { $packageName = "Microsoft.ML.OnnxRuntime.Gpu.Windows" }
+          "cuda" { $packageName = if ($arch -eq "arm64") { "Microsoft.ML.OnnxRuntime" } else { "Microsoft.ML.OnnxRuntime.Gpu.Windows" } }
           "directml" { $packageName = "Microsoft.ML.OnnxRuntime.DirectML" }
           default { $packageName = "Microsoft.ML.OnnxRuntime" }
       }
@@ -123,6 +175,13 @@ steps:
       $feedProject = "2692857e-05ef-43b4-ba9c-ccf1c22c437c"
       $feedId = "7982ae20-ed19-4a35-a362-a96ac99897b7"
       $version = "${{ parameters.ort_build_version }}"
+      if ($version -eq "default") {
+          $version = "$(ort_version)"
+      }
+      if ([string]::IsNullOrWhiteSpace($version)) {
+          Write-Error "ORT build version is empty. Set ort_build_version or pass the appropriate ORT version parameter for $ep/$arch."
+          exit 1
+      }
 
       $source = "${{ parameters.ort_build_source }}"
       if ($source -eq "Ort-Nightly") {
@@ -148,7 +207,8 @@ steps:
       # Copy headers
       # CUDA packages store headers in buildTransitive/native/include, while other
       # ORT packages (CPU, DML) use build/native/include per NuGet package layout.
-      if ($ep -eq "cuda") {
+      # CUDA arm64 currently builds against the CPU ORT package, which uses build/native/include.
+      if ($ep -eq "cuda" -and $arch -ne "arm64") {
           $headerSrc = "$extractPath\buildTransitive\native\include"
       } else {
           $headerSrc = "$extractPath\build\native\include"
@@ -162,13 +222,13 @@ steps:
       Write-Host "ORT_HOME: $ortHome"
       Get-ChildItem "$ortHome" -Recurse | Select-Object FullName
       Write-Host "##vso[task.setvariable variable=ORT_HOME_CMAKE_ARG]-DORT_HOME=$ortHome"
-    displayName: 'Download ORT build package ${{ parameters.ort_build_version }}'
+    displayName: 'Download ORT build package'
     condition: ne(variables['ep'], 'winml')
     workingDirectory: '$(Build.Repository.LocalPath)'
 
 - powershell: |
-    cmake --preset windows_$(arch)_$(ep)_$(build_config) -DWINML_SDK_VERSION='$(ort_version)'
-  displayName: 'Configure CMake C API with WinML (arm64, NO CUDA) - $(ort_version)'
+    cmake --preset windows_$(arch)_$(ep)_$(build_config) -T cuda='$(Build.Repository.LocalPath)\cuda_sdk\v$(cuda_version)' -DWINML_SDK_VERSION='$(ort_version)' -DCMAKE_CUDA_ARCHITECTURES='$(cuda_archs)'
+  displayName: 'Configure CMake C API with WinML (arm64, WITH CUDA) - $(ort_version)'
   condition: and(eq(variables['ep'], 'winml'), eq(variables['arch'], 'arm64'))
   workingDirectory: '$(Build.Repository.LocalPath)'
 

.pipelines/stages/nuget-packaging-stage.yml

@@ -105,6 +105,7 @@ stages:
           build_config: ${{ parameters.build_config }}
           enable_linux_cuda: ${{ parameters.enable_linux_cuda }}
           enable_win_cuda: ${{ parameters.enable_win_cuda }}
+          enable_win_arm64: ${{ parameters.enable_win_arm64 }}
           nuget_version_type: ${{ parameters.nuget_version_type }}
     - ${{ if eq(parameters.enable_win_dml, true) }}:
       - template: jobs/nuget-packaging-job.yml