Merge branch 'main' into update-cuda-13.1.1

Author: leofang

.gitattributes

@@ -6,7 +6,9 @@ cuda/_version.py export-subst
 # we do not own any headers checked in, don't touch them
 *.h binary
 *.hpp binary
-# Exception: headers we own (cuda_core C++ implementation)
+# Exception: headers we own
+cuda_bindings/cuda/bindings/_bindings/*.h -binary text diff
+cuda_bindings/cuda/bindings/_lib/*.h -binary text diff
 cuda_core/cuda/core/_cpp/*.h -binary text diff
 cuda_core/cuda/core/_cpp/*.hpp -binary text diff
 # git should not convert line endings in PNG files

.github/actions/fetch_ctk/action.yml

@@ -14,7 +14,7 @@ inputs:
   cuda-components:
     description: "A list of the CTK components to install as a comma-separated list. e.g. 'cuda_nvcc,cuda_nvrtc,cuda_cudart'"
     required: false
-    default: "cuda_nvcc,cuda_cudart,cuda_crt,libnvvm,cuda_nvrtc,cuda_profiler_api,cuda_cccl,libnvjitlink,libcufile"
+    default: "cuda_nvcc,cuda_cudart,cuda_crt,libnvvm,cuda_nvrtc,cuda_profiler_api,cuda_cccl,libnvjitlink,libcufile,libnvfatbin"
   cuda-path:
     description: "where the CTK components will be installed to, relative to $PWD"
     required: false

.github/workflows/backport.yml

@@ -9,14 +9,25 @@ on:
     types: [closed, labeled]
     branches:
       - main
+  workflow_dispatch:
+    inputs:
+      backport-branch:
+        description: "Branch to backport commits onto"
+        required: false
+        type: string
+      pull-request:
+        description: "PR to backport"
+        required: true
+        type: number
+
 
 permissions:
   contents: write       # so it can comment
   pull-requests: write  # so it can create pull requests
 
 jobs:
-  backport:
-    name: Backport pull request
+  backport-from-pr:
+    name: Backport directly from a pull request
     if: ${{ github.repository_owner == 'nvidia' &&
             github.event.pull_request.merged == true &&
             contains( github.event.pull_request.labels.*.name, 'to-be-backported')
@@ -38,3 +49,28 @@ jobs:
           copy_labels_pattern: true
           copy_requested_reviewers: true
           target_branches: ${{ env.OLD_BRANCH }}
+  backport-to-branch:
+    name: Backport a specific PR against a specific branch
+    if: github.repository_owner == 'nvidia' && github.event_name == 'workflow_dispatch'
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@8e8c483db84b4bee98b60c0593521ed34d9990e8  # v6.0.1
+
+      - name: Load branch from environment name
+        if: inputs.backport-branch == null
+        run: |
+          BACKPORT_BRANCH=$(yq '.backport_branch' ci/versions.yml)
+          echo "BACKPORT_BRANCH=${BRANCH}" >> $GITHUB_ENV
+
+      - name: Load branch name from input
+        if: inputs.backport-branch != null
+        run: echo "BACKPORT_BRANCH=${{ inputs.backport-branch }}" >> $GITHUB_ENV
+
+      - name: Create backport pull requests
+        uses: korthout/backport-action@c656f5d5851037b2b38fb5db2691a03fa229e3b2  # v4.0.1
+        with:
+          copy_assignees: true
+          copy_labels_pattern: true
+          copy_requested_reviewers: true
+          target_branches: ${{ env.BACKPORT_BRANCH }}
+          source_pr_number: ${{ inputs.pull-request }}

.github/workflows/build-docs.yml

@@ -47,6 +47,12 @@ jobs:
       run:
         shell: bash -el {0}
     steps:
+      - name: validate build-ctk
+        run: |
+          if [ ! "${{ inputs.build-ctk-ver }}" =~ ^[0-9]+\.[0-9]+\.[0-9]+$ ]; then
+            echo "error: `build-ctk-ver` ${{ inputs.build-ctk-ver }} version does not match MAJOR.MINOR.MICRO" >&2
+            exit 1
+          fi
       - name: Checkout ${{ github.event.repository.name }}
         uses: actions/checkout@8e8c483db84b4bee98b60c0593521ed34d9990e8  # v6.0.1
         with:

.github/workflows/ci.yml

@@ -40,6 +40,7 @@ jobs:
     runs-on: ubuntu-latest
     outputs:
       skip: ${{ steps.get-should-skip.outputs.skip }}
+      doc-only: ${{ steps.get-should-skip.outputs.doc_only }}
     steps:
       - name: Checkout repository
         uses: actions/checkout@8e8c483db84b4bee98b60c0593521ed34d9990e8  # v6.0.1
@@ -50,13 +51,23 @@ jobs:
         run: |
           set -euxo pipefail
           if ${{ startsWith(github.ref_name, 'pull-request/') }}; then
-            skip="$(gh pr view "$(grep -Po '(\d+)$' <<< '${{ github.ref_name }}')" --json title --jq '.title | contains("[no-ci]")')"
+            pr_number="$(grep -Po '(\d+)$' <<< '${{ github.ref_name }}')"
+            pr_title="$(gh pr view "${pr_number}" --json title --jq '.title')"
+            skip="$(echo "${pr_title}" | grep -q '\[no-ci\]' && echo true || echo false)"
+            doc_only="$(echo "${pr_title}" | grep -q '\[doc-only\]' && echo true || echo false)"
           else
             skip=false
+            doc_only=false
           fi
           echo "skip=${skip}" >> "$GITHUB_OUTPUT"
+          echo "doc_only=${doc_only}" >> "$GITHUB_OUTPUT"
 
-  # WARNING: make sure all of the build jobs are in sync
+  # NOTE: Build jobs are intentionally split by platform rather than using a single
+  # matrix. This allows each test job to depend only on its corresponding build,
+  # so faster platforms can proceed through build & test without waiting for slower
+  # ones. Keep these job definitions textually identical except for:
+  #   - host-platform value
+  #   - if: condition (build-linux-64 omits doc-only check since it's needed for docs)
   build-linux-64:
     needs:
       - ci-vars
@@ -75,7 +86,7 @@ jobs:
       cuda-version: ${{ needs.ci-vars.outputs.CUDA_BUILD_VER }}
       prev-cuda-version: ${{ needs.ci-vars.outputs.CUDA_PREV_BUILD_VER }}
 
-  # WARNING: make sure all of the build jobs are in sync
+  # See build-linux-64 for why build jobs are split by platform.
   build-linux-aarch64:
     needs:
       - ci-vars
@@ -86,15 +97,15 @@ jobs:
         host-platform:
           - linux-aarch64
     name: Build ${{ matrix.host-platform }}, CUDA ${{ needs.ci-vars.outputs.CUDA_BUILD_VER }}
-    if: ${{ github.repository_owner == 'nvidia' && !fromJSON(needs.should-skip.outputs.skip) }}
+    if: ${{ github.repository_owner == 'nvidia' && !fromJSON(needs.should-skip.outputs.skip) && !fromJSON(needs.should-skip.outputs.doc-only) }}
     secrets: inherit
     uses: ./.github/workflows/build-wheel.yml
     with:
       host-platform: ${{ matrix.host-platform }}
       cuda-version: ${{ needs.ci-vars.outputs.CUDA_BUILD_VER }}
       prev-cuda-version: ${{ needs.ci-vars.outputs.CUDA_PREV_BUILD_VER }}
 
-  # WARNING: make sure all of the build jobs are in sync
+  # See build-linux-64 for why build jobs are split by platform.
   build-windows:
     needs:
       - ci-vars
@@ -105,27 +116,32 @@ jobs:
         host-platform:
           - win-64
     name: Build ${{ matrix.host-platform }}, CUDA ${{ needs.ci-vars.outputs.CUDA_BUILD_VER }}
-    if: ${{ github.repository_owner == 'nvidia' && !fromJSON(needs.should-skip.outputs.skip) }}
+    if: ${{ github.repository_owner == 'nvidia' && !fromJSON(needs.should-skip.outputs.skip) && !fromJSON(needs.should-skip.outputs.doc-only) }}
     secrets: inherit
     uses: ./.github/workflows/build-wheel.yml
     with:
       host-platform: ${{ matrix.host-platform }}
       cuda-version: ${{ needs.ci-vars.outputs.CUDA_BUILD_VER }}
       prev-cuda-version: ${{ needs.ci-vars.outputs.CUDA_PREV_BUILD_VER }}
 
-  # WARNING: make sure both Linux test jobs are in sync
+  # NOTE: Test jobs are split by platform for the same reason as build jobs (see
+  # build-linux-64). Keep these job definitions textually identical except for:
+  #   - host-platform value
+  #   - build job under needs:
+  #   - uses: (test-wheel-linux.yml vs test-wheel-windows.yml)
   test-linux-64:
     strategy:
       fail-fast: false
       matrix:
         host-platform:
           - linux-64
     name: Test ${{ matrix.host-platform }}
-    if: ${{ github.repository_owner == 'nvidia' }}
+    if: ${{ github.repository_owner == 'nvidia' && !fromJSON(needs.should-skip.outputs.doc-only) }}
     permissions:
       contents: read  # This is required for actions/checkout
     needs:
       - ci-vars
+      - should-skip
       - build-linux-64
     secrets: inherit
     uses: ./.github/workflows/test-wheel-linux.yml
@@ -134,19 +150,20 @@ jobs:
       host-platform: ${{ matrix.host-platform }}
       build-ctk-ver: ${{ needs.ci-vars.outputs.CUDA_BUILD_VER }}
 
-  # WARNING: make sure both Linux test jobs are in sync
+  # See test-linux-64 for why test jobs are split by platform.
   test-linux-aarch64:
     strategy:
       fail-fast: false
       matrix:
         host-platform:
           - linux-aarch64
     name: Test ${{ matrix.host-platform }}
-    if: ${{ github.repository_owner == 'nvidia' }}
+    if: ${{ github.repository_owner == 'nvidia' && !fromJSON(needs.should-skip.outputs.doc-only) }}
     permissions:
       contents: read  # This is required for actions/checkout
     needs:
       - ci-vars
+      - should-skip
       - build-linux-aarch64
     secrets: inherit
     uses: ./.github/workflows/test-wheel-linux.yml
@@ -155,18 +172,20 @@ jobs:
       host-platform: ${{ matrix.host-platform }}
       build-ctk-ver: ${{ needs.ci-vars.outputs.CUDA_BUILD_VER }}
 
+  # See test-linux-64 for why test jobs are split by platform.
   test-windows:
     strategy:
       fail-fast: false
       matrix:
         host-platform:
           - win-64
     name: Test ${{ matrix.host-platform }}
-    if: ${{ github.repository_owner == 'nvidia' }}
+    if: ${{ github.repository_owner == 'nvidia' && !fromJSON(needs.should-skip.outputs.doc-only) }}
     permissions:
       contents: read  # This is required for actions/checkout
     needs:
       - ci-vars
+      - should-skip
       - build-windows
     secrets: inherit
     uses: ./.github/workflows/test-wheel-windows.yml
@@ -196,6 +215,7 @@ jobs:
     if: always()
     runs-on: ubuntu-latest
     needs:
+      - should-skip
       - test-linux-64
       - test-linux-aarch64
       - test-windows
@@ -219,11 +239,18 @@ jobs:
           # failing job(s) will timeout causing a cancellation here and the
           # build to succeed which we don't want (originally this was just
           # 'exit 0')
-          if ${{ needs.test-linux-64.result == 'cancelled' ||
-                 needs.test-linux-aarch64.result == 'cancelled' ||
-                 needs.test-windows.result == 'cancelled' ||
-                 needs.doc.result == 'cancelled' }}; then
+          #
+          # Note: When [doc-only] is in PR title, test jobs are intentionally
+          # skipped and should not cause failure.
+          doc_only=${{ needs.should-skip.outputs.doc-only }}
+          if ${{ needs.doc.result == 'cancelled' }}; then
             exit 1
-          else
-            exit 0
           fi
+          if [[ "${doc_only}" != "true" ]]; then
+            if ${{ needs.test-linux-64.result == 'cancelled' ||
+                   needs.test-linux-aarch64.result == 'cancelled' ||
+                   needs.test-windows.result == 'cancelled' }}; then
+              exit 1
+            fi
+          fi
+          exit 0

.github/workflows/release.yml

@@ -135,6 +135,7 @@ jobs:
     needs:
       - check-tag
       - determine-run-id
+      - doc
     secrets: inherit
     uses: ./.github/workflows/release-upload.yml
     with:
@@ -148,6 +149,7 @@ jobs:
     needs:
       - check-tag
       - determine-run-id
+      - doc
     environment:
       name: ${{ inputs.wheel-dst }}
       url: https://${{ (inputs.wheel-dst == 'testpypi' && 'test.') || '' }}pypi.org/p/${{ inputs.component }}/

.github/workflows/test-wheel-linux.yml

@@ -57,12 +57,6 @@ jobs:
           # Read base matrix from YAML file for the specific architecture
           TEST_MATRIX=$(yq -o json ".linux[\"${MATRIX_TYPE}\"] | map(select(.ARCH == \"${ARCH}\"))" ci/test-matrix.yml)
 
-          # Add special runner for amd64 if applicable
-          if [[ "${ARCH}" == "amd64" ]]; then
-            SPECIAL_RUNNERS=$(yq -o json '.linux.special_runners.amd64' ci/test-matrix.yml)
-            TEST_MATRIX=$(echo "$TEST_MATRIX" | yq -o json ". + $SPECIAL_RUNNERS")
-          fi
-
           # Apply matrix filter and wrap in include structure
           MATRIX=$(echo "$TEST_MATRIX" | jq -c '${{ inputs.matrix_filter }} | if (. | length) > 0 then {include: .} else "Error: Empty matrix\n" | halt_error(1) end')
 
@@ -73,12 +67,12 @@ jobs:
           echo "OLD_BRANCH=${OLD_BRANCH}" >> "$GITHUB_OUTPUT"
 
   test:
-    name: py${{ matrix.PY_VER }}, ${{ matrix.CUDA_VER }}, ${{ (matrix.LOCAL_CTK == '1' && 'local') || 'wheels' }}, ${{ matrix.GPU }}
+    name: py${{ matrix.PY_VER }}, ${{ matrix.CUDA_VER }}, ${{ (matrix.LOCAL_CTK == '1' && 'local') || 'wheels' }}, ${{ matrix.GPU }}${{ matrix.GPU_COUNT != '1' && format('(x{0})', matrix.GPU_COUNT) || '' }}
     needs: compute-matrix
     strategy:
       fail-fast: false
       matrix: ${{ fromJSON(needs.compute-matrix.outputs.MATRIX) }}
-    runs-on: "linux-${{ matrix.ARCH }}-gpu-${{ matrix.GPU }}-${{ matrix.DRIVER }}-1"
+    runs-on: "linux-${{ matrix.ARCH }}-gpu-${{ matrix.GPU }}-${{ matrix.DRIVER }}-${{ matrix.GPU_COUNT }}"
     # The build stage could fail but we want the CI to keep moving.
     if: ${{ github.repository_owner == 'nvidia' && !cancelled() }}
     # Our self-hosted runners require a container
@@ -269,17 +263,15 @@ jobs:
       - name: Ensure cuda-python installable
         run: |
           if [[ "${{ matrix.LOCAL_CTK }}" == 1 ]]; then
-            pip install cuda_python*.whl
+            pip install --only-binary=:all: cuda_python*.whl
           else
-            pip install $(ls cuda_python*.whl)[all]
+            pip install --only-binary=:all: $(ls cuda_python*.whl)[all]
           fi
 
       - name: Install cuda.pathfinder extra wheels for testing
         run: |
           set -euo pipefail
           pushd cuda_pathfinder
-          # Install pathfinder from the pre-built wheel, since building from
-          # source won't work in this context (we don't have a git checkout).
           pip install --only-binary=:all: -v ./*.whl --group "test-cu${TEST_CUDA_MAJOR}"
           pip list
           popd

.github/workflows/test-wheel-windows.yml

@@ -66,7 +66,7 @@ jobs:
       fail-fast: false
       matrix: ${{ fromJSON(needs.compute-matrix.outputs.MATRIX) }}
     if: ${{ github.repository_owner == 'nvidia' && !cancelled() }}
-    runs-on: "windows-${{ matrix.ARCH }}-gpu-${{ matrix.GPU }}-${{ matrix.DRIVER }}-1"
+    runs-on: "windows-${{ matrix.ARCH }}-gpu-${{ matrix.GPU }}-${{ matrix.DRIVER }}-${{ matrix.GPU_COUNT }}"
     steps:
       - name: Checkout ${{ github.event.repository.name }}
         uses: actions/checkout@8e8c483db84b4bee98b60c0593521ed34d9990e8  # v6.0.1
@@ -236,16 +236,16 @@ jobs:
       - name: Ensure cuda-python installable
         run: |
           if ('${{ matrix.LOCAL_CTK }}' -eq '1') {
-            pip install (Get-ChildItem -Filter cuda_python*.whl).FullName
+            pip install --only-binary=:all: (Get-ChildItem -Filter cuda_python*.whl).FullName
           } else {
-            pip install "$((Get-ChildItem -Filter cuda_python*.whl).FullName)[all]"
+            pip install --only-binary=:all: "$((Get-ChildItem -Filter cuda_python*.whl).FullName)[all]"
           }
 
       - name: Install cuda.pathfinder extra wheels for testing
         shell: bash --noprofile --norc -xeuo pipefail {0}
         run: |
           pushd cuda_pathfinder
-          pip install --only-binary=:all: -v . --group "test-cu${TEST_CUDA_MAJOR}"
+          pip install --only-binary=:all: -v ./*.whl --group "test-cu${TEST_CUDA_MAJOR}"
           pip list
           popd
 

.pre-commit-config.yaml

@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-FileCopyrightText: Copyright (c) 2024-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 #
 # SPDX-License-Identifier: Apache-2.0
 
@@ -30,6 +30,7 @@ repos:
         additional_dependencies:
           - https://files.pythonhosted.org/packages/cc/20/ff623b09d963f88bfde16306a54e12ee5ea43e9b597108672ff3a408aad6/pathspec-0.12.1-py3-none-any.whl
         exclude: '.*pixi\.lock'
+        args: ["--fix"]
 
       - id: no-markdown-in-docs-source
         name: Prevent markdown files in docs/source directories
@@ -89,5 +90,6 @@ repos:
         args: [--no-pycodestyle]
         exclude: ^cuda_bindings/
 
+
 default_language_version:
       python: python3

README.md

@@ -13,6 +13,7 @@ CUDA Python is the home for accessing NVIDIA’s CUDA platform from Python. It c
 * [nvshmem4py](https://docs.nvidia.com/nvshmem/api/api/language_bindings/python/index.html): Pythonic interface to the NVSHMEM library, enabling Python applications to leverage NVSHMEM's high-performance PGAS (Partitioned Global Address Space) programming model for GPU-accelerated computing
 * [Nsight Python](https://docs.nvidia.com/nsight-python/index.html): Python kernel profiling interface that automates performance analysis across multiple kernel configurations using NVIDIA Nsight Tools
 * [CUPTI Python](https://docs.nvidia.com/cupti-python/): Python APIs for creation of profiling tools that target CUDA Python applications via the CUDA Profiling Tools Interface (CUPTI)
+* [Accelerated Computing Hub](https://github.com/NVIDIA/accelerated-computing-hub): Open-source learning materials related to GPU computing. You will find user guides, tutorials, and other works freely available for all learners interested in GPU computing.
 
 CUDA Python is currently undergoing an overhaul to improve existing and introduce new components. All of the previously available functionality from the `cuda-python` package will continue to be available, please refer to the [cuda.bindings](https://nvidia.github.io/cuda-python/cuda-bindings/latest) documentation for installation guide and further detail.