pytorch · lanluo-nvidia · Jul 30, 2026 · Jul 30, 2026 · Jul 30, 2026 · Jul 30, 2026
diff --git a/.github/workflows/build_windows_arm64.yml b/.github/workflows/build_windows_arm64.yml
@@ -0,0 +1,88 @@
+name: Build Windows ARM64 RTX wheel
+
+on:
+  workflow_call:
+    inputs:
+      runner:
+        description: Windows x86_64 runner with the local ARM64 PyTorch artifact
+        required: true
+        type: string
+      python-version:
+        required: true
+        type: string
+      desired-cuda:
+        description: cu130 or cu132
+        required: true
+        type: string
+      pytorch-arm64-artifact:
+        description: Absolute runner-local path to the Windows ARM64 PyTorch wheel or zip
+        required: true
+        type: string
+      tensorrt-rtx-arm64-artifact:
+        description: Runner-local path or URL for the Windows ARM64 TensorRT-RTX SDK zip
+        default: https://developer.nvidia.com/downloads/trt/rtx_sdk/secure/1.6/TensorRT-RTX-1.6.1.120-Windows-arm64-cuda-13.4-Release-external.zip
+        type: string
+
+permissions:
+  contents: read
+
+jobs:
+  build:
+    runs-on: ${{ inputs.runner }}
+    timeout-minutes: 120
+    env:
+      CU_VERSION: ${{ inputs.desired-cuda }}
+      CHANNEL: nightly
+      USE_TRT_RTX: "1"
+      PYTHON_VERSION: ${{ inputs.python-version }}
+    defaults:
+      run:
+        shell: bash -l {0}
+    steps:
+      - uses: actions/checkout@v6
+        with:
+          submodules: recursive
+      - uses: actions/checkout@v6
+        with:
+          repository: pytorch/test-infra
+          path: test-infra
+      - uses: ./test-infra/.github/actions/setup-windows
+        with:
+          cuda-version: ${{ inputs.desired-cuda }}
+      - uses: actions/setup-python@v6
+        with:
+          python-version: ${{ inputs.python-version }}
+          architecture: x64
+      - name: Install x64 host build dependencies
+        shell: pwsh
+        run: |
+          python -m pip install --upgrade pip
+          python -m pip install --pre torch --index-url "https://download.pytorch.org/whl/nightly/${{ inputs.desired-cuda }}"
+          python -m pip install numpy packaging pyyaml setuptools wheel fmt
+      - name: Prepare ARM64 target sysroot
+        shell: pwsh
+        run: |
+          ./packaging/prepare_windows_arm64_cross.ps1 `
+            -PyTorchArtifact '${{ inputs.pytorch-arm64-artifact }}' `
+            -TensorRTArtifact '${{ inputs.tensorrt-rtx-arm64-artifact }}' `
+            -TargetRoot '${{ runner.temp }}\torchtrt-windows-arm64' `
+            -CudaRoot "$env:CUDA_PATH"
+      - name: Prepare Torch-TensorRT dependencies
+        run: bash packaging/pre_build_script_windows.sh
+      - name: Cross-compile Windows ARM64 wheel
+        shell: cmd
+        run: |
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" amd64_arm64
+          call packaging\build_windows_arm64_cross.bat
+      - name: Validate Windows ARM64 wheel
+        shell: cmd
+        run: |
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" amd64_arm64
+          powershell -NoProfile -ExecutionPolicy Bypass -File packaging\validate_windows_arm64_wheel.ps1 -DistDirectory dist
+      - uses: actions/upload-artifact@v6
+        with:
+          name: torch-tensorrt-rtx-${{ inputs.python-version }}-${{ inputs.desired-cuda }}-win-arm64
+          path: dist/*-win_arm64.whl
+          if-no-files-found: error
+      - uses: ./test-infra/.github/actions/teardown-windows
+        if: always()
diff --git a/.github/workflows/ci-windows-arm64.yml b/.github/workflows/ci-windows-arm64.yml
@@ -0,0 +1,99 @@
+name: CI Windows ARM64
+
+# RTX-only build channel. Until Windows ARM64 PyTorch is published remotely,
+# this is manually dispatched onto an x86_64 Windows runner that can access the
+# two runner-local PyTorch artifacts. Runtime tests begin when a native ARM64
+# runner is available; cross-builds perform PE/header validation only.
+on:
+  workflow_dispatch:
+    inputs:
+      runner:
+        description: Windows x86_64 runner label with access to the local artifacts
+        required: true
+        type: string
+      python-version:
+        description: Target CPython version matching both ARM64 PyTorch wheels
+        default: "3.12"
+        required: true
+        type: string
+      cuda:
+        description: CUDA variants to build
+        default: both
+        type: choice
+        options: [both, cu130, cu132]
+      pytorch-cu130-artifact:
+        description: Absolute runner-local path to the cu130 Windows ARM64 PyTorch artifact
+        required: false
+        type: string
+      pytorch-cu132-artifact:
+        description: Absolute runner-local path to the cu132 Windows ARM64 PyTorch artifact
+        required: false
+        type: string
+      tensorrt-rtx-artifact:
+        description: Optional local TensorRT-RTX ARM64 zip; blank uses the NVIDIA URL
+        required: false
+        type: string
+
+permissions:
+  contents: read
+
+jobs:
+  validate-inputs:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Validate selected artifacts
+        env:
+          CUDA_SELECTION: ${{ inputs.cuda }}
+          PYTORCH_CU130: ${{ inputs.pytorch-cu130-artifact }}
+          PYTORCH_CU132: ${{ inputs.pytorch-cu132-artifact }}
+        run: |
+          set -euo pipefail
+          if [[ "${CUDA_SELECTION}" != "cu132" && -z "${PYTORCH_CU130}" ]]; then
+            echo "::error::pytorch-cu130-artifact is required for ${CUDA_SELECTION}"
+            exit 1
+          fi
+          if [[ "${CUDA_SELECTION}" != "cu130" && -z "${PYTORCH_CU132}" ]]; then
+            echo "::error::pytorch-cu132-artifact is required for ${CUDA_SELECTION}"
+            exit 1
+          fi
+
+  cu130:
+    needs: validate-inputs
+    if: inputs.cuda == 'both' || inputs.cuda == 'cu130'
+    uses: ./.github/workflows/build_windows_arm64.yml
+    with:
+      runner: ${{ inputs.runner }}
+      python-version: ${{ inputs.python-version }}
+      desired-cuda: cu130
+      pytorch-arm64-artifact: ${{ inputs.pytorch-cu130-artifact }}
+      tensorrt-rtx-arm64-artifact: ${{ inputs.tensorrt-rtx-artifact || 'https://developer.nvidia.com/downloads/trt/rtx_sdk/secure/1.6/TensorRT-RTX-1.6.1.120-Windows-arm64-cuda-13.4-Release-external.zip' }}
+
+  cu132:
+    needs: validate-inputs
+    if: inputs.cuda == 'both' || inputs.cuda == 'cu132'
+    uses: ./.github/workflows/build_windows_arm64.yml
+    with:
+      runner: ${{ inputs.runner }}
+      python-version: ${{ inputs.python-version }}
+      desired-cuda: cu132
+      pytorch-arm64-artifact: ${{ inputs.pytorch-cu132-artifact }}
+      tensorrt-rtx-arm64-artifact: ${{ inputs.tensorrt-rtx-artifact || 'https://developer.nvidia.com/downloads/trt/rtx_sdk/secure/1.6/TensorRT-RTX-1.6.1.120-Windows-arm64-cuda-13.4-Release-external.zip' }}
+
+  gate:
+    needs: [validate-inputs, cu130, cu132]
+    if: always()
+    runs-on: ubuntu-latest
+    steps:
+      - name: Gate Windows ARM64 builds
+        env:
+          VALIDATE_RESULT: ${{ needs.validate-inputs.result }}
+          CU130_RESULT: ${{ needs.cu130.result }}
+          CU132_RESULT: ${{ needs.cu132.result }}
+        run: |
+          set -euo pipefail
+          for result in "${VALIDATE_RESULT}" "${CU130_RESULT}" "${CU132_RESULT}"; do
+            if [[ "${result}" == "failure" || "${result}" == "cancelled" ]]; then
+              exit 1
+            fi
+          done
+          echo "Windows ARM64 RTX cross-build gate passed."
diff --git a/BUILD.bazel b/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",
     }),

diff --git a/MODULE.bazel b/MODULE.bazel
@@ -11,6 +11,9 @@ bazel_dep(name = "rules_python", version = "1.3.0")
 bazel_dep(name = "bazel_skylib", version = "1.7.1")
 bazel_dep(name = "rules_foreign_cc", version = "0.15.1")
 
+cc_configure = use_extension("@rules_cc//cc:extensions.bzl", "cc_configure_extension")
+use_repo(cc_configure, "local_config_cc")
+
 python = use_extension("@rules_python//python/extensions:python.bzl", "python")
 python.toolchain(
     ignore_root_user_error = True,
@@ -94,6 +97,13 @@ new_local_repository(
     path = "C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v13.2/",
 )
 
+# supposedly cuda 13.4 on windows x86 preview has everything required for cross compile for windows on arm
+new_local_repository(
+    name = "cuda_win_arm64",
+    build_file = "@//third_party/cuda:BUILD",
+    path = "C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v13.4/",
+)
+
 http_archive = use_repo_rule("@bazel_tools//tools/build_defs/repo:http.bzl", "http_archive")
 
 ####################################################################################
@@ -163,6 +173,10 @@ http_archive(
     urls = ["https://download.pytorch.org/libtorch/nightly/cu130/libtorch-win-shared-with-deps-latest.zip"],
 )
 
+# Native Windows ARM64 builds use the Torch package installed in the active
+# Python environment. Cross-builds override discovery through TORCH_PATH.
+local_torch(name = "libtorch_win_arm64")
+
 http_archive(
     name = "torch_l4t",
     build_file = "@//third_party/libtorch:BUILD",
@@ -241,3 +255,12 @@ http_archive(
         "https://developer.nvidia.com/downloads/trt/rtx_sdk/secure/1.6/TensorRT-RTX-1.6.1.120-Windows-amd64-cuda-13.4-Release-external.zip",
     ],
 )
+
+http_archive(
+    name = "tensorrt_rtx_win_arm64",
+    build_file = "@//third_party/tensorrt_rtx/archive:BUILD",
+    strip_prefix = "TensorRT-RTX-1.6.1.120",
+    urls = [
+        "https://developer.nvidia.com/downloads/trt/rtx_sdk/secure/1.6/TensorRT-RTX-1.6.1.120-Windows-arm64-cuda-13.4-Release-external.zip",
+    ],
+)
diff --git a/core/BUILD b/core/BUILD
@@ -1,6 +1,7 @@
 load("@rules_cc//cc:defs.bzl", "cc_library")
 load("@rules_pkg//:pkg.bzl", "pkg_tar")
 load("@rules_pkg//pkg:mappings.bzl", "pkg_files")
+
 package(default_visibility = ["//visibility:public"])
 
 config_setting(
@@ -33,8 +34,20 @@ 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 = {
@@ -92,15 +105,17 @@ cc_library(
         "//core/util/logging",
     ] + select({
         ":jetpack": ["@tensorrt_l4t//:nvinfer"],
-        ":rtx_win": ["@tensorrt_rtx_win//:nvinfer"],
-        ":rtx_x86_64": ["@tensorrt_rtx//:nvinfer"],
         ":rtx_sbsa": ["@tensorrt_rtx_sbsa//:nvinfer"],
+        ":rtx_win_arm64": ["@tensorrt_rtx_win_arm64//:nvinfer"],
+        ":rtx_win_x86_64": ["@tensorrt_rtx_win//:nvinfer"],
+        ":rtx_x86_64": ["@tensorrt_rtx//:nvinfer"],
         ":sbsa": ["@tensorrt_sbsa//:nvinfer"],
         ":windows": ["@tensorrt_win//:nvinfer"],
         "//conditions:default": ["@tensorrt//:nvinfer"],
     }) + select({
         ":jetpack": ["@torch_l4t//:libtorch"],
-        ":rtx_win": ["@libtorch_win//:libtorch"],
+        ":rtx_win_arm64": ["@libtorch_win_arm64//:libtorch"],
+        ":rtx_win_x86_64": ["@libtorch_win//:libtorch"],
         ":use_torch_whl": ["@torch_whl//:libtorch"],
         ":windows": ["@libtorch_win//:libtorch"],
         "//conditions:default": ["@libtorch"],
@@ -121,6 +136,6 @@ pkg_files(
     srcs = [
         "compiler.h",
     ],
-    visibility = ["//visibility:public"],
     prefix = "include/torch_tensorrt/core/",
+    visibility = ["//visibility:public"],
 )