windows on arm build

.github/workflows/build_windows.yml

@@ -105,6 +105,14 @@ on:
         required: false
         default: false
         type: boolean
+      cross-compile:
+        description: 'Build ARM64 from an x64 Windows host'
+        default: false
+        type: boolean
+      skip-binary-upload:
+        description: 'Skip the test-infra binary upload job'
+        default: false
+        type: boolean
     secrets:
       R2_ACCOUNT_ID:
         description: Cloudflare R2 account ID
@@ -186,7 +194,7 @@ jobs:
           setup-miniconda: false
           python-version: ${{ env.PYTHON_VERSION }}
           cuda-version: ${{ env.CU_VERSION }}
-          arch: ${{ inputs.architecture }}
+          arch: ${{ inputs.cross-compile && 'arm64' || inputs.architecture }}
       - name: Shorten Conda environment path
         if: inputs.architecture == 'x64'
         run: |
@@ -233,6 +241,22 @@ jobs:
           source "${BUILD_ENV_FILE}"
           # shellcheck disable=SC2086
           ${CONDA_RUN} ${PIP_INSTALL_TORCH}
+      - name: Install CUDA 13.4 toolkit with ARM64 target libraries
+        if: inputs.cross-compile
+        shell: powershell
+        env:
+          CUDA_INSTALLER_URL: https://packages.nvidia.com/prerelease/cuda/13.4.0/local_installers/cuda_13.4.0_windows_x86_64.exe
+          CUDA_ROOT: C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4
+        run: |
+          $installer = Join-Path $env:RUNNER_TEMP "cuda_13.4.0_windows_x86_64.exe"
+          Invoke-WebRequest -Uri $env:CUDA_INSTALLER_URL -OutFile $installer
+          $process = Start-Process -FilePath $installer -ArgumentList "-s" -Wait -PassThru
+          if ($process.ExitCode -ne 0) { throw "CUDA 13.4 installer failed with exit code $($process.ExitCode)" }
+          $cudaArm64Lib = Join-Path $env:CUDA_ROOT "lib\arm64\cudart.lib"
+          if (-not (Test-Path -LiteralPath $cudaArm64Lib -PathType Leaf)) { throw "CUDA 13.4 ARM64 target library was not installed: $cudaArm64Lib" }
+          "CUDA_PATH=$env:CUDA_ROOT" | Out-File -FilePath $env:GITHUB_ENV -Encoding utf8 -Append
+          "CUDA_HOME=$env:CUDA_ROOT" | Out-File -FilePath $env:GITHUB_ENV -Encoding utf8 -Append
+          (Join-Path $env:CUDA_ROOT "bin") | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
       - name: Run Pre-Script with Caching
         if: ${{ inputs.pre-script != '' && inputs.architecture == 'x64' }}
         uses: ./test-infra/.github/actions/run-script-with-cache
@@ -242,6 +266,58 @@ jobs:
           repository: ${{ inputs.repository  }}
           script: ${{ inputs.pre-script }}
           is_windows: 'enabled'
+      - name: Prepare ARM64 target sysroot
+        if: inputs.cross-compile
+        working-directory: ${{ inputs.repository }}
+        shell: powershell
+        run: |
+          ./packaging/prepare_windows_arm64_cross.ps1 `
+            -TargetRoot '${{ runner.temp }}\torchtrt-windows-arm64' `
+            -CudaRoot $env:CUDA_PATH `
+            -PythonExecutable (Join-Path $env:CONDA_ENV 'python.exe')
+      - name: Ensure MSVC ARM64 cross tools
+        if: inputs.cross-compile
+        shell: powershell
+        run: |
+          $vswhere = Join-Path ${env:ProgramFiles(x86)} "Microsoft Visual Studio/Installer/vswhere.exe"
+          if (-not (Test-Path -LiteralPath $vswhere -PathType Leaf)) { throw "Visual Studio locator not found: $vswhere" }
+          $vsInstall = (& $vswhere -latest -products * -requires Microsoft.VisualStudio.Component.VC.Tools.x86.x64 -property installationPath | Select-Object -First 1).Trim()
+          if ([string]::IsNullOrWhiteSpace($vsInstall)) { throw "No Visual Studio instance with x64 MSVC tools was found" }
+          $vcvarsall = [IO.Path]::Combine($vsInstall, "VC", "Auxiliary", "Build", "vcvarsall.bat")
+          if (-not (Test-Path -LiteralPath $vcvarsall -PathType Leaf)) { throw "vcvarsall.bat not found: $vcvarsall" }
+          $arm64Compiler = Get-ChildItem -Path (Join-Path $vsInstall "VC/Tools/MSVC/*/bin/Hostx64/arm64/cl.exe") -File -ErrorAction SilentlyContinue | Select-Object -First 1
+          if (-not $arm64Compiler) {
+            $installStopwatch = [Diagnostics.Stopwatch]::StartNew()
+            $setup = Join-Path ${env:ProgramFiles(x86)} "Microsoft Visual Studio/Installer/setup.exe"
+            if (-not (Test-Path -LiteralPath $setup -PathType Leaf)) { throw "Visual Studio Installer not found: $setup" }
+            & $setup modify --installPath $vsInstall --add Microsoft.VisualStudio.Component.VC.Tools.ARM64 --passive --norestart
+            if ($LASTEXITCODE -ne 0) { throw "Installing the MSVC ARM64 component failed with exit code $LASTEXITCODE" }
+            $deadline = (Get-Date).AddMinutes(20)
+            while (-not $arm64Compiler -and (Get-Date) -lt $deadline) {
+              Write-Host "Waiting for the MSVC ARM64 compiler to be installed ($([Math]::Floor($installStopwatch.Elapsed.TotalSeconds)) seconds elapsed; next check in 30 seconds)..."
+              Start-Sleep -Seconds 30
+              $arm64Compiler = Get-ChildItem -Path (Join-Path $vsInstall "VC/Tools/MSVC/*/bin/Hostx64/arm64/cl.exe") -File -ErrorAction SilentlyContinue | Select-Object -First 1
+            }
+            $installStopwatch.Stop()
+            if ($arm64Compiler) { Write-Host "MSVC ARM64 compiler became available after $([Math]::Round($installStopwatch.Elapsed.TotalSeconds, 1)) seconds" }
+            $arm64Compiler = Get-ChildItem -Path (Join-Path $vsInstall "VC/Tools/MSVC/*/bin/Hostx64/arm64/cl.exe") -File -ErrorAction SilentlyContinue | Select-Object -First 1
+          }
+          if (-not $arm64Compiler) { throw "MSVC ARM64 compiler was not installed after $([Math]::Round($installStopwatch.Elapsed.TotalSeconds, 1)) seconds" }
+          "TORCHTRT_VCVARSALL=$vcvarsall" | Out-File -FilePath $env:GITHUB_ENV -Encoding utf8 -Append
+      - name: Cross-compile Windows ARM64 wheel
+        if: inputs.cross-compile
+        working-directory: ${{ inputs.repository }}
+        shell: cmd
+        run: |
+          call "%TORCHTRT_VCVARSALL%" amd64_arm64
+          %CONDA_RUN% cmd /c packaging\build_windows_arm64_cross.bat
+      - name: Validate Windows ARM64 wheel
+        if: inputs.cross-compile
+        working-directory: ${{ inputs.repository }}
+        shell: cmd
+        run: |
+          call "%TORCHTRT_VCVARSALL%" amd64_arm64
+          powershell -NoProfile -ExecutionPolicy Bypass -File packaging\validate_windows_arm64_wheel.ps1 -DistDirectory dist
       - name: Run Pre-Script Arm64
         if: ${{ inputs.pre-script != '' && inputs.architecture == 'arm64' }}
         env:
@@ -278,7 +354,7 @@ jobs:
       - name: Build clean
         # Only relevant for setup-py builds; python-build-package projects may
         # not have a setup.py (e.g. scikit-build-core based builds).
-        if: inputs.build-platform == 'setup-py' && inputs.architecture == 'x64'
+        if: inputs.build-platform == 'setup-py' && inputs.architecture == 'x64' && inputs.cross-compile == false
         working-directory: ${{ inputs.repository }}
         env:
           ENV_SCRIPT: ${{ inputs.env-script }}
@@ -295,7 +371,7 @@ jobs:
             fi
           fi
       - name: Set PYTORCH_VERSION on x64
-        if: inputs.architecture == 'x64'
+        if: inputs.architecture == 'x64' && inputs.cross-compile == false
         working-directory: ${{ inputs.repository }}
         run: |
           source "${BUILD_ENV_FILE}"
@@ -307,7 +383,7 @@ jobs:
             export PYTORCH_VERSION="$(${CONDA_RUN} pip show torch | grep ^Version: | sed 's/Version:  *//')"
           fi
       - name: Build the wheel (python-build-package) X64
-        if: ${{ inputs.build-platform == 'python-build-package' && inputs.architecture == 'x64' }}
+        if: ${{ inputs.build-platform == 'python-build-package' && inputs.architecture == 'x64' && inputs.cross-compile == false }}
         working-directory: ${{ inputs.repository }}
         env:
           ENV_SCRIPT: ${{ inputs.env-script }}
@@ -322,7 +398,7 @@ jobs:
             ${CONDA_RUN} ${ENV_SCRIPT} ${{ inputs.build-command }}
           fi
       - name: Build the wheel (setup-py) X64
-        if: ${{ inputs.build-platform == 'setup-py' && inputs.architecture == 'x64' }}
+        if: ${{ inputs.build-platform == 'setup-py' && inputs.architecture == 'x64' && inputs.cross-compile == false }}
         working-directory: ${{ inputs.repository }}
         env:
           ENV_SCRIPT: ${{ inputs.env-script }}
@@ -398,13 +474,13 @@ jobs:
         env:
           POST_SCRIPT: ${{ inputs.post-script }}
           ENV_SCRIPT: ${{ inputs.env-script }}
-        if: ${{ inputs.post-script != '' && inputs.architecture == 'x64'}}
+        if: ${{ inputs.post-script != '' && inputs.architecture == 'x64' && inputs.cross-compile == false }}
         run: |
           set -euxo pipefail
           source "${BUILD_ENV_FILE}"
           ${CONDA_RUN} ${ENV_SCRIPT} ${POST_SCRIPT}
       - name: Smoke Test X64
-        if: ${{ inputs.architecture == 'x64' && inputs.is-release-tarball == false }}
+        if: ${{ inputs.architecture == 'x64' && inputs.cross-compile == false && inputs.is-release-tarball == false }}
         env:
           ENV_SCRIPT: ${{ inputs.env-script }}
           PACKAGE_NAME: ${{ inputs.package-name }}
@@ -454,15 +530,20 @@ jobs:
           fi
       # NB: Only upload to GitHub after passing smoke tests
       - name: Get Artifact name
-        if: inputs.architecture == 'arm64'
+        if: inputs.architecture == 'arm64' || inputs.cross-compile
         env:
           REPOSITORY: ${{ inputs.repository }}
           REF: ${{ inputs.ref }}
           PYTHON_VERSION: ${{ matrix.python_version }}
           CU_VERSION: ${{ env.CU_VERSION }}
-          ARCH: ${{ inputs.architecture }}
+          ARCH: ${{ inputs.cross-compile && 'arm64' || inputs.architecture }}
         run: |
-          echo "ARTIFACT_NAME=${REPOSITORY//\//_}_${REF//\//_}_${PYTHON_VERSION}_${CU_VERSION}_${ARCH}" >> "${GITHUB_ENV}"
+          artifact_cuda_version="${CU_VERSION}"
+          # The cross-compiled ARM64 sysroot is CUDA 13.4; retain the host CUDA label otherwise.
+          if [[ "${{ inputs.cross-compile }}" == "true" ]]; then
+            artifact_cuda_version="cu134"
+          fi
+          echo "ARTIFACT_NAME=${REPOSITORY//\//_}_${REF//\//_}_${PYTHON_VERSION}_${artifact_cuda_version}_${ARCH}" >> "${GITHUB_ENV}"
       - name: Upload wheel to GitHub
         if: ${{ inputs.is-release-tarball == false }}
         continue-on-error: true
@@ -478,14 +559,14 @@ jobs:
           name: cxx11-zip-${{ env.PYTHON_VERSION }}-${{ env.CU_VERSION }}
           path: ${{ inputs.repository }}/release/tarball/
       - uses: ./test-infra/.github/actions/teardown-windows
-        if: inputs.architecture == 'x64'
+        if: inputs.architecture == 'x64' || inputs.cross-compile
         name: Teardown Windows
 
   upload:
     needs: build
     uses: pytorch/test-infra/.github/workflows/_binary_upload.yml@main
     # if it is the release wheel or tarball, then skip upload to pytorch index(release wheel are uploaded to pypi)
-    if: ${{ inputs.is-release-wheel == false && inputs.is-release-tarball == false }}
+    if: ${{ inputs.skip-binary-upload == false && inputs.is-release-wheel == false && inputs.is-release-tarball == false }}
     with:
       repository: ${{ inputs.repository }}
       ref: ${{ inputs.ref }}

.github/workflows/ci-linux-x86_64.yml

@@ -5,8 +5,8 @@ name: CI Linux x86_64
 # come from the shared _decide reusable.
 
 on:
-  pull_request:
-    types: [opened, synchronize, reopened, labeled]
+  #pull_request:
+    #types: [opened, synchronize, reopened, labeled]
   pull_request_review:
     types: [submitted]
   push:

.github/workflows/ci-sbsa.yml

@@ -7,8 +7,8 @@ name: CI SBSA
 # RTX channels build @tensorrt_rtx_sbsa via USE_TRT_RTX=1.)
 
 on:
-  pull_request:
-    types: [opened, synchronize, reopened, labeled]
+  #pull_request:
+    #types: [opened, synchronize, reopened, labeled]
   pull_request_review:
     types: [submitted]
   push:

.github/workflows/ci-windows-arm64.yml

@@ -0,0 +1,53 @@
+name: CI Windows ARM64
+
+on:
+  pull_request:
+    types: [opened, synchronize, reopened, labeled]
+  pull_request_review:
+    types: [submitted]
+  push:
+    branches: [main]
+  schedule:
+    - cron: "0 7 * * *"
+  workflow_dispatch:
+concurrency:
+  # Keep one active ARM64 run per PR/ref; CI-control labels supersede the prior run,
+  # while unrelated label churn remains isolated.
+  group: "${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}${{ (github.event.action == 'labeled' || github.event.action == 'unlabeled') && !contains(fromJSON('[\"ci: full\", \"ci: nightly\", \"backend: TensorRT\", \"backend: TensorRT-RTX\"]'), github.event.label.name) && format('-label-{0}', github.event.label.name) || '' }}"
+  cancel-in-progress: true
+
+
+permissions:
+  id-token: write
+  contents: read
+
+jobs:
+
+  build:
+    uses: ./.github/workflows/build_windows.yml
+    with:
+      repository: ${{ github.repository }}
+      ref: ${{ github.sha }}
+      build-matrix: >-
+        {"include":[{"build_name":"windows-arm64-cross","python_version":"3.13","desired_cuda":"cu132","validation_runner":"windows.g5.4xlarge.nvidia.gpu"}]}
+      architecture: x64
+      cross-compile: true
+      pre-script: packaging/pre_build_script_windows.sh
+      use-rtx: true
+      timeout: 180
+      skip-binary-upload: true
+
+  gate:
+    needs: [build]
+    if: always()
+    runs-on: ubuntu-latest
+    steps:
+      - name: Gate Windows ARM64 builds
+        env:
+          BUILD_RESULT: ${{ needs.build.result }}
+        run: |
+          set -euo pipefail
+          if [[ "${BUILD_RESULT}" != "success" ]]; then
+            exit 1
+          fi
+          echo "Windows ARM64 RTX cross-build gate passed."

BUILD.bazel

@@ -1,5 +1,5 @@
-load("@hedron_compile_commands//:refresh_compile_commands.bzl", "refresh_compile_commands")
 load("@bazel_skylib//rules:common_settings.bzl", "bool_flag")
+load("@hedron_compile_commands//:refresh_compile_commands.bzl", "refresh_compile_commands")
 load("@rules_pkg//:pkg.bzl", "pkg_tar", "pkg_zip")
 load("@rules_pkg//pkg:mappings.bzl", "pkg_files")
 
@@ -16,8 +16,20 @@ config_setting(
 )
 
 config_setting(
-    name = "rtx_win",
+    name = "rtx_win_x86_64",
+    constraint_values = [
+        "@platforms//cpu:x86_64",
+        "@platforms//os:windows",
+    ],
+    flag_values = {
+        "//toolchains/dep_collection:compute_libs": "rtx",
+    },
+)
+
+config_setting(
+    name = "rtx_win_arm64",
     constraint_values = [
+        "@platforms//cpu:arm64",
         "@platforms//os:windows",
     ],
     flag_values = {
@@ -226,9 +238,10 @@ pkg_tar(
     srcs = ["//cpp/bin/torchtrtc"] + select({
         ":jetpack": [],
         ":no_package_executorch": [],
-        ":rtx_win": [],
-        ":rtx_x86_64": [],
         ":rtx_sbsa": [],
+        ":rtx_win_arm64": [],
+        ":rtx_win_x86_64": [],
+        ":rtx_x86_64": [],
         ":windows": [],
         "//conditions:default": [
             "//examples/executorch_reference_runner:example_executorch_runner",
@@ -270,16 +283,18 @@ pkg_tar(
     ] + select({
         ":jetpack": [],
         ":no_package_executorch": [],
-        ":rtx_win": [],
-        ":rtx_x86_64": [],
         ":rtx_sbsa": [],
+        ":rtx_win_arm64": [],
+        ":rtx_win_x86_64": [],
+        ":rtx_x86_64": [],
         ":windows": [],
         "//conditions:default": [
             ":executorch_source_package",
             ":include_executorch",
         ],
     }) + select({
-        ":rtx_win": [],
+        ":rtx_win_arm64": [],
+        ":rtx_win_x86_64": [],
         ":windows": [],
         "//conditions:default": [":bin"],
     }),
@@ -300,7 +315,8 @@ pkg_tar(
 alias(
     name = "libtorchtrt",
     actual = select({
-        ":rtx_win": ":libtorchtrt_zip",
+        ":rtx_win_arm64": ":libtorchtrt_zip",
+        ":rtx_win_x86_64": ":libtorchtrt_zip",
         ":windows": ":libtorchtrt_zip",
         "//conditions:default": ":libtorchtrt_tar",
     }),
@@ -326,7 +342,8 @@ pkg_tar(
 alias(
     name = "libtorchtrt_runtime",
     actual = select({
-        ":rtx_win": ":libtorchtrt_runtime_zip",
+        ":rtx_win_arm64": ":libtorchtrt_runtime_zip",
+        ":rtx_win_x86_64": ":libtorchtrt_runtime_zip",
         ":windows": ":libtorchtrt_runtime_zip",
         "//conditions:default": ":libtorchtrt_runtime_tar",
     }),