Pull 13.4.1 changes into main

.github/RELEASE-core.md

@@ -89,21 +89,6 @@ changes, so they are only permitted at a major-version boundary per the
 
 ---
 
-## Refresh frozen fallback explanation tables
-
-When this release adds support for a new CUDA Toolkit version, check
-whether the CTK release added any new `CUresult` or `cudaError_t` codes.
-If so, update the frozen fallback tables so error messages stay
-informative for consumers on older `cuda-bindings` versions:
-
-- `cuda_core/cuda/core/_utils/driver_cu_result_explanations_frozen.py`
-- `cuda_core/cuda/core/_utils/runtime_cuda_error_explanations_frozen.py`
-
-The corresponding tests in `test_utils_enum_explanations_helpers.py`
-will fail if an entry is missing.
-
----
-
 ## Finalize the doc update, including release notes
 
 Review every PR included in the release. For each one, check whether new

.github/actions/fetch_ctk/action.yml

@@ -79,7 +79,7 @@ runs:
           function extract() {
             tar -xvf $1 -C $CACHE_TMP_DIR --strip-components=1
           }
-        elif [[ "${{ inputs.host-platform }}" == "win-64" ]]; then
+        elif [[ "${{ inputs.host-platform }}" == win* ]]; then
           function extract() {
             _TEMP_DIR_=$(mktemp -d)
             unzip $1 -d $_TEMP_DIR_
@@ -120,8 +120,12 @@ runs:
         # Note: try to escape | and > ...
         tar -czvf ${CTK_CACHE_FILENAME} ${CACHE_TMP_DIR}
 
-        # "Move" files from temp dir to CUDA_PATH
-        CUDA_PATH="./cuda_toolkit"
+        # Populate the final CUDA_PATH directly (never stage through ./cuda_toolkit
+        # unconditionally — a second call with a different cuda-path would leave
+        # symlinks from a prerelease CTK at ./cuda_toolkit/include etc., causing
+        # "cp: cannot overwrite non-directory" failures on the next restore).
+        CUDA_PATH="${{ inputs.cuda-path }}"
+        rm -rf $CUDA_PATH
         mkdir -p $CUDA_PATH
         # Unfortunately we cannot use "rsync -av $CACHE_TMP_DIR/ $CUDA_PATH" because
         # not all runners have rsync pre-installed (or even installable, such as
@@ -144,22 +148,14 @@ runs:
       run: |
         ls -l
         CACHE_TMP_DIR="./cache_tmp_dir"
-        CUDA_PATH="./cuda_toolkit"
+        CUDA_PATH="${{ inputs.cuda-path }}"
+        rm -rf $CUDA_PATH
         mkdir -p $CUDA_PATH
         tar -xzvf $CTK_CACHE_FILENAME
         # Can't use rsync here, see above
         cp -r $CACHE_TMP_DIR/* $CUDA_PATH
         rm -rf $CACHE_TMP_DIR $CTK_CACHE_FILENAME
         ls -l $CUDA_PATH
-        if [ ! -d "$CUDA_PATH/include" ]; then
-          exit 1
-        fi
-
-    - name: Move CTK to the specified location
-      if: ${{ inputs.cuda-path != './cuda_toolkit' }}
-      shell: bash --noprofile --norc -xeuo pipefail {0}
-      run: |
-        mv ./cuda_toolkit ${{ inputs.cuda-path }}
 
     - name: Set output environment variables
       shell: bash --noprofile --norc -xeuo pipefail {0}

.github/workflows/build-wheel.yml

@@ -19,6 +19,11 @@ on:
         required: false
         type: string
         default: ""
+      single-cuda-major:
+        description: "Build wheels for only the current CUDA major; skip the prior-major build and wheel merge"
+        required: false
+        type: boolean
+        default: false
 
 defaults:
   run:
@@ -51,10 +56,16 @@ jobs:
           - "3.14t"
           - "3.15"
           - "3.15t"
+        exclude:
+          # CPython 3.10 has no official Windows ARM64 build (neither
+          # nuget-cpython nor actions/setup-python's manifest carries one),
+          # so it cannot be built or tested on win-arm64.
+          - python-version: ${{ (inputs.host-platform == 'win-arm64' && '3.10') || '' }}
     name: py${{ matrix.python-version }}
     runs-on: ${{ (inputs.host-platform == 'linux-64' && 'linux-amd64-cpu8') ||
                  (inputs.host-platform == 'linux-aarch64' && 'linux-arm64-cpu8') ||
-                 (inputs.host-platform == 'win-64' && 'windows-2022') }}
+                 (inputs.host-platform == 'win-64' && 'windows-2022') ||
+                 (inputs.host-platform == 'win-arm64' && 'windows-11-arm') }}
     steps:
       - name: Checkout ${{ github.event.repository.name }}
         uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
@@ -87,7 +98,7 @@ jobs:
         uses: nv-gha-runners/setup-proxy-cache@main
         continue-on-error: true
         # Skip cache on GitHub-hosted Windows runners.
-        if: ${{ inputs.host-platform != 'win-64' }}
+        if: ${{ !startsWith(inputs.host-platform, 'win') }}
         with:
           enable-apt: true
 
@@ -98,22 +109,26 @@ jobs:
           # WAR: setup-python is not relocatable, and cibuildwheel hard-wires to 3.12...
           # see https://github.com/actions/setup-python/issues/871
           python-version: "3.12"
+          architecture: ${{ ((inputs.host-platform == 'linux-aarch64' || inputs.host-platform == 'win-arm64') && 'arm64') || 'x64' }}
 
       - name: Set up MSVC
         if: ${{ startsWith(inputs.host-platform, 'win') && (env.BUILD_BINDINGS == 'true' || env.BUILD_CORE == 'true' || env.TEST_BINDINGS == 'true' || env.TEST_CORE == 'true') }}
         uses: step-security/msvc-dev-cmd@22c98154b708dbd743e6f27a933cf6ceba3305c4  # v1.13.1
+        with:
+          arch: ${{ (inputs.host-platform == 'win-arm64' && 'arm64') || 'x64' }}
 
       - name: Set up yq
         # GitHub made an unprofessional decision to not provide it in their Windows VMs,
         # see https://github.com/actions/runner-images/issues/7443.
         if: ${{ startsWith(inputs.host-platform, 'win') && env.BUILD_CORE == 'true' }}
         env:
           YQ_VERSION: v4.52.5
-          YQ_SHA256: 47594981f3848a4b4447494adeca9555f908f7cf0a89c4da3fd0243a4631da1c
+          YQ_ARCH: ${{ (inputs.host-platform == 'win-arm64' && 'arm64') || 'amd64' }}
+          YQ_SHA256: ${{ (inputs.host-platform == 'win-arm64' && '236867affa7f18701d4c763cf16b6df962cf4f7e89a8570a5954cf94a38f41c7') || '47594981f3848a4b4447494adeca9555f908f7cf0a89c4da3fd0243a4631da1c' }}
           YQ_DIR: yq
         shell: pwsh -command ". '{0}'"
         run: |
-          $yqUrl = "https://github.com/mikefarah/yq/releases/download/${env:YQ_VERSION}/yq_windows_amd64.exe"
+          $yqUrl = "https://github.com/mikefarah/yq/releases/download/${env:YQ_VERSION}/yq_windows_${env:YQ_ARCH}.exe"
           mkdir -Force -ErrorAction SilentlyContinue "${env:YQ_DIR}" | Out-Null
           Invoke-WebRequest -UseBasicParsing -OutFile "${env:YQ_DIR}/yq.exe" -Uri "$yqUrl"
           $hash = (Get-FileHash -Algorithm SHA256 "${env:YQ_DIR}/yq.exe").Hash.ToLower()
@@ -253,7 +268,7 @@ jobs:
             echo "ok!"
 
       - name: Report sccache stats (cuda.bindings)
-        if: ${{ env.BUILD_BINDINGS == 'true' && inputs.host-platform != 'win-64' }}
+        if: ${{ env.BUILD_BINDINGS == 'true' && !startsWith(inputs.host-platform, 'win') }}
         uses: ./.github/actions/sccache-summary
         with:
           json-file: sccache_bindings.json
@@ -362,7 +377,7 @@ jobs:
             echo "ok!"
 
       - name: Report sccache stats (cuda.core)
-        if: ${{ env.BUILD_CORE == 'true' && inputs.host-platform != 'win-64' }}
+        if: ${{ env.BUILD_CORE == 'true' && !startsWith(inputs.host-platform, 'win') }}
         uses: ./.github/actions/sccache-summary
         with:
           json-file: sccache_core.json
@@ -524,7 +539,7 @@ jobs:
 
       # Note: This overwrites CUDA_PATH etc
       - name: Set up mini CTK
-        if: ${{ env.BUILD_CORE == 'true' || env.TEST_CORE == 'true' }}
+        if: ${{ !inputs.single-cuda-major && (env.BUILD_CORE == 'true' || env.TEST_CORE == 'true') }}
         uses: ./.github/actions/fetch_ctk
         continue-on-error: false
         with:
@@ -533,13 +548,13 @@ jobs:
           cuda-path: "./cuda_toolkit_prev"
 
       - name: Build cuda.core test binaries
-        if: ${{ env.TEST_CORE == 'true' }}
+        if: ${{ !inputs.single-cuda-major && env.TEST_CORE == 'true' }}
         run: |
           nvcc --version
           python "${{ env.CUDA_CORE_TEST_BINARIES_DIR }}/build_test_binaries.py"
 
       - name: Upload cuda.core test binaries
-        if: ${{ env.TEST_CORE == 'true' }}
+        if: ${{ !inputs.single-cuda-major && env.TEST_CORE == 'true' }}
         uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a  # v7.0.1
         with:
           name: ${{ env.CUDA_CORE_ARTIFACT_NAME }}-test-binaries
@@ -550,7 +565,7 @@ jobs:
           if-no-files-found: error
 
       - name: Download cuda.bindings build artifacts from the prior branch
-        if: ${{ env.BUILD_CORE == 'true' }}
+        if: ${{ !inputs.single-cuda-major && env.BUILD_CORE == 'true' }}
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
         run: |
@@ -586,7 +601,7 @@ jobs:
           rmdir "${OLD_ARTIFACT_DIR}"
 
       - name: Constrain previous cuda.core to the downloaded cuda.bindings wheel
-        if: ${{ env.BUILD_CORE == 'true' }}
+        if: ${{ !inputs.single-cuda-major && env.BUILD_CORE == 'true' }}
         run: |
           pathfinder_wheels=(cuda_pathfinder/cuda_pathfinder-*.whl)
           bindings_wheels=(cuda_bindings/dist-prev/cuda_bindings-"${BUILD_PREV_CUDA_MAJOR}".*.whl)
@@ -608,7 +623,7 @@ jobs:
           } | tee wheel-constraints/cuda-core-prev.txt
 
       - name: Build cuda.core wheel
-        if: ${{ env.BUILD_CORE == 'true' }}
+        if: ${{ !inputs.single-cuda-major && env.BUILD_CORE == 'true' }}
         uses: pypa/cibuildwheel@1828c10ab37f080699c7b81cea34097c684a7074  # v4.2.0
         with:
           package-dir: ./cuda_core/
@@ -656,15 +671,15 @@ jobs:
             echo "ok!"
 
       - name: Report sccache stats (cuda.core prev)
-        if: ${{ env.BUILD_CORE == 'true' && inputs.host-platform != 'win-64' }}
+        if: ${{ !inputs.single-cuda-major && env.BUILD_CORE == 'true' && !startsWith(inputs.host-platform, 'win') }}
         uses: ./.github/actions/sccache-summary
         with:
           json-file: sccache_core_prev.json
           label: "cuda.core (prev CTK)"
           build-step: "Build cuda.core wheel"
 
       - name: List the cuda.core artifacts directory and rename
-        if: ${{ env.BUILD_CORE == 'true' }}
+        if: ${{ !inputs.single-cuda-major && env.BUILD_CORE == 'true' }}
         run: |
           if [[ "${{ inputs.host-platform }}" == win* ]]; then
             export CHOWN=chown
@@ -688,14 +703,23 @@ jobs:
           ls -lahR ${{ env.CUDA_CORE_ARTIFACTS_DIR }}
 
       - name: Merge cuda.core wheels
-        if: ${{ env.BUILD_CORE == 'true' }}
+        if: ${{ !inputs.single-cuda-major && env.BUILD_CORE == 'true' }}
         run: |
           pip install wheel
           python ci/tools/merge_cuda_core_wheels.py \
             "${{ env.CUDA_CORE_ARTIFACTS_DIR }}"/cu"${BUILD_CUDA_MAJOR}"/cuda_core*.whl \
             "${{ env.CUDA_CORE_ARTIFACTS_DIR }}"/cu"${BUILD_PREV_CUDA_MAJOR}"/cuda_core*.whl \
             --output-dir "${{ env.CUDA_CORE_ARTIFACTS_DIR }}"
 
+      - name: Finalize single-major cuda.core wheel
+        if: ${{ inputs.single-cuda-major && env.BUILD_CORE == 'true' }}
+        run: |
+          for wheel in "${{ env.CUDA_CORE_ARTIFACTS_DIR }}"/cu"${BUILD_CUDA_MAJOR}"/*.cu"${BUILD_CUDA_MAJOR}".whl; do
+            base_name=$(basename "${wheel}" ".cu${BUILD_CUDA_MAJOR}.whl")
+            mv "${wheel}" "${{ env.CUDA_CORE_ARTIFACTS_DIR }}/${base_name}.whl"
+          done
+          ls -lahR "${{ env.CUDA_CORE_ARTIFACTS_DIR }}"
+
       - name: Check cuda.core wheel
         if: ${{ env.BUILD_CORE == 'true' }}
         run: |

.github/workflows/ci.yml

@@ -33,6 +33,8 @@ jobs:
     outputs:
       CUDA_BUILD_VER: ${{ steps.get-vars.outputs.cuda_build_ver }}
       CUDA_PREV_BUILD_VER: ${{ steps.get-vars.outputs.cuda_prev_build_ver }}
+      WINDOWS_ARM64_SUPPORTED: ${{ steps.get-vars.outputs.windows_arm64_supported }}
+      WINDOWS_ARM64_SINGLE_CUDA_MAJOR: ${{ steps.get-vars.outputs.windows_arm64_single_cuda_major }}
     steps:
       - name: Checkout repository
         uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
@@ -47,6 +49,33 @@ jobs:
           cuda_prev_build_ver=$(yq '.cuda.prev_build.version' ci/versions.yml)
           echo "cuda_prev_build_ver=$cuda_prev_build_ver" >> $GITHUB_OUTPUT
 
+          # Windows ARM64 is available starting with CUDA 13.4.
+          if [[ "$cuda_build_ver" =~ ^([0-9]+)\.([0-9]+)(\.|$) ]]; then
+            cuda_build_major="${BASH_REMATCH[1]}"
+            cuda_build_minor="${BASH_REMATCH[2]}"
+          else
+            echo "Invalid CUDA build version: $cuda_build_ver" >&2
+            exit 1
+          fi
+          if (( cuda_build_major > 13 || (cuda_build_major == 13 && cuda_build_minor >= 4) )); then
+            windows_arm64_supported=true
+          else
+            windows_arm64_supported=false
+          fi
+          echo "windows_arm64_supported=$windows_arm64_supported" >> $GITHUB_OUTPUT
+
+          # No CUDA 13 windows-arm64 toolkit exists for a major other than the
+          # current one (windows-arm64 support started mid-way through the 13.x
+          # series), so cuda.core can only be built against a single CUDA major
+          # while the build major is still 13. Once the build major advances to
+          # 14, a CUDA 13 windows-arm64 toolkit will exist as the prior major.
+          if [[ "$cuda_build_major" == "13" ]]; then
+            windows_arm64_single_cuda_major=true
+          else
+            windows_arm64_single_cuda_major=false
+          fi
+          echo "windows_arm64_single_cuda_major=$windows_arm64_single_cuda_major" >> $GITHUB_OUTPUT
+
   should-skip:
     if: ${{ github.repository_owner == 'nvidia' }}
     runs-on: ubuntu-latest
@@ -414,6 +443,31 @@ jobs:
       prev-cuda-version: ${{ needs.ci-vars.outputs.CUDA_PREV_BUILD_VER }}
       workplan: ${{ needs.detect-changes.outputs.workplan }}
 
+  # Windows ARM64 is available starting with CUDA 13.4. Build only the current
+  # CUDA major because no prior-major toolkit or wheel artifacts exist.
+  build-windows-arm64:
+    needs:
+      - ci-vars
+      - should-skip
+      - detect-changes
+    name: Build win-arm64, CUDA ${{ needs.ci-vars.outputs.CUDA_BUILD_VER }}
+    if: ${{ github.repository_owner == 'nvidia' &&
+            !fromJSON(needs.should-skip.outputs.skip) &&
+            !fromJSON(needs.should-skip.outputs.doc-only) &&
+            needs.ci-vars.outputs.WINDOWS_ARM64_SUPPORTED == 'true' &&
+            fromJSON(needs.detect-changes.outputs.workplan).jobs.platforms.windows }}
+    permissions:
+      actions: read
+      contents: read
+    secrets: inherit
+    uses: ./.github/workflows/build-wheel.yml
+    with:
+      host-platform: win-arm64
+      cuda-version: ${{ needs.ci-vars.outputs.CUDA_BUILD_VER }}
+      prev-cuda-version: ${{ needs.ci-vars.outputs.CUDA_PREV_BUILD_VER }}
+      workplan: ${{ needs.detect-changes.outputs.workplan }}
+      single-cuda-major: ${{ needs.ci-vars.outputs.WINDOWS_ARM64_SINGLE_CUDA_MAJOR == 'true' }}
+
   # NOTE: test-sdist jobs are split by platform (mirroring build-* and test-wheel-*)
   # so platform-specific sources (e.g. cuda_bindings/*_windows.pyx selected by
   # build_hooks.py) are exercised on their target OS. Keep these job definitions
@@ -613,6 +667,7 @@ jobs:
       - build-linux-64
       - build-linux-aarch64
       - build-windows
+      - build-windows-arm64
       - test-sdist-linux
       - test-sdist-windows
       - test-linux-64
@@ -644,6 +699,7 @@ jobs:
           doc_only="${{ needs.should-skip.outputs.doc-only }}"
           linux_selected="${{ needs.detect-changes.outputs.workplan && fromJSON(needs.detect-changes.outputs.workplan).jobs.platforms.linux || false }}"
           windows_selected="${{ needs.detect-changes.outputs.workplan && fromJSON(needs.detect-changes.outputs.workplan).jobs.platforms.windows || false }}"
+          windows_arm64_supported="${{ needs.ci-vars.outputs.WINDOWS_ARM64_SUPPORTED }}"
           build_selected="${{ needs.detect-changes.outputs.workplan && fromJSON(needs.detect-changes.outputs.workplan).jobs.sdist_tests || false }}"
           run_core_api_check="${{ needs.detect-changes.outputs.workplan && fromJSON(needs.detect-changes.outputs.workplan).jobs.core_api_checks || false }}"
           is_pr="${{ startsWith(github.ref_name, 'pull-request/') }}"
@@ -679,6 +735,12 @@ jobs:
           check_result "build-linux-aarch64" "$linux_expected"
           check_result "build-windows"       "$windows_expected"
 
+          windows_arm64_expected="skipped"
+          if [[ "$windows_arm64_supported" == "true" ]]; then
+            windows_arm64_expected="$windows_expected"
+          fi
+          check_result "build-windows-arm64" "$windows_arm64_expected"
+
           # Sdist tests follow build selection; wheel tests follow the platform plan.
           expected="skipped"
           if [[ "$doc_only" != "true" && "$build_selected" == "true" ]]; then

.pre-commit-config.yaml

@@ -103,6 +103,7 @@ repos:
           - --max-retries=3
           - --no-progress
         files: '\.(md|rst)$'
+        exclude: ^qa/
 
   # Standard hooks
   - repo: https://github.com/pre-commit/pre-commit-hooks

CONTRIBUTING.md

@@ -48,7 +48,7 @@ correct.** Each package matches its own tag prefix:
 
 | Package | Tag pattern |
 | --- | --- |
-| `cuda-bindings`, `cuda-python` | `v*` (e.g. `v13.3.1`) |
+| `cuda-bindings`, `cuda-python` | `v*` (e.g. `v13.4.1`) |
 | `cuda-core` | `cuda-core-v*` (e.g. `cuda-core-v1.1.0`) |
 | `cuda-pathfinder` | `cuda-pathfinder-v*` (e.g. `cuda-pathfinder-v1.6.0`) |
 
@@ -109,7 +109,7 @@ version-check failure:
    `cuda-bindings` to that same bogus version.
 2. **Stale tags** (a fork that has not fetched upstream in a while): you get a
    plausible-looking but wrong version, e.g. `13.0.4.dev650+g0d22cb44` when the
-   real latest tag is `v13.3.1`. Nothing warns you. Note there is no leading
+   real latest tag is `v13.4.1`. Nothing warns you. Note there is no leading
    `v` — the tag prefix is stripped by `tag_regex`.
 3. **No git metadata** (source zip): the build fails with
    `LookupError: setuptools-scm was unable to detect version`.