Set up a new VM-based CI infrastructure

.github/actions/fetch_ctk/action.yml

@@ -0,0 +1,193 @@
+# SPDX-FileCopyrightText: Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+#
+# SPDX-License-Identifier: Apache-2.0
+
+name: Fetch mini CTK
+
+description: Fetch (or create) a mini CUDA Toolkit from cache
+
+inputs:
+  host-platform:
+    required: true
+    type: string
+  cuda-version:
+    required: true
+    type: string
+  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
+    type: string
+    default: "cuda_nvcc,cuda_cudart,cuda_crt,libnvvm,cuda_nvrtc,cuda_profiler_api,cuda_cccl,libnvjitlink,libcufile"
+  cuda-path:
+    description: "where the CTK components will be installed to, relative to $PWD"
+    required: false
+    type: string
+    default: "./cuda_toolkit"
+
+runs:
+  using: composite
+  steps:
+    - name: Set up CTK cache variable
+      shell: bash --noprofile --norc -xeuo pipefail {0}
+      run: |
+        # Pre-process the component list to ensure hash uniqueness
+        CTK_CACHE_COMPONENTS=${{ inputs.cuda-components }}
+        # Conditionally strip out libnvjitlink for CUDA versions < 12
+        CUDA_MAJOR_VER="$(cut -d '.' -f 1 <<< ${{ inputs.cuda-version }})"
+        if [[ "$CUDA_MAJOR_VER" -lt 12 ]]; then
+          CTK_CACHE_COMPONENTS="${CTK_CACHE_COMPONENTS//libnvjitlink/}"
+        fi
+        # Conditionally strip out cuda_crt and libnvvm for CUDA versions < 13
+        CUDA_MAJOR_VER="$(cut -d '.' -f 1 <<< ${{ inputs.cuda-version }})"
+        if [[ "$CUDA_MAJOR_VER" -lt 13 ]]; then
+          CTK_CACHE_COMPONENTS="${CTK_CACHE_COMPONENTS//cuda_crt/}"
+          CTK_CACHE_COMPONENTS="${CTK_CACHE_COMPONENTS//libnvvm/}"
+        fi
+        # Conditionally strip out libcufile since it does not support Windows
+        if [[ "${{ inputs.host-platform }}" == win-* ]]; then
+          CTK_CACHE_COMPONENTS="${CTK_CACHE_COMPONENTS//libcufile/}"
+        fi
+        # Cleanup stray commas after removing components
+        CTK_CACHE_COMPONENTS="${CTK_CACHE_COMPONENTS//,,/,}"
+
+        HASH=$(echo -n "${CTK_CACHE_COMPONENTS}" | sha256sum | awk '{print $1}')
+        echo "CTK_CACHE_KEY=mini-ctk-${{ inputs.cuda-version }}-${{ inputs.host-platform }}-$HASH" >> $GITHUB_ENV
+        echo "CTK_CACHE_FILENAME=mini-ctk-${{ inputs.cuda-version }}-${{ inputs.host-platform }}-$HASH.tar.gz" >> $GITHUB_ENV
+        echo "CTK_CACHE_COMPONENTS=${CTK_CACHE_COMPONENTS}" >> $GITHUB_ENV
+
+    - name: Install dependencies
+      uses: ./.github/actions/install_unix_deps
+      continue-on-error: false
+      with:
+        dependencies: "zstd curl xz-utils"
+        dependent_exes: "zstd curl xz"
+
+    - name: Download CTK cache
+      id: ctk-get-cache
+      uses: actions/cache/restore@5a3ec84eff668545956fd18022155c47e93e2684  # v4.2.3
+      continue-on-error: true
+      with:
+        key: ${{ env.CTK_CACHE_KEY }}
+        path: ./${{ env.CTK_CACHE_FILENAME }}
+        fail-on-cache-miss: false
+
+    - name: Get CUDA components
+      if: ${{ steps.ctk-get-cache.outputs.cache-hit != 'true' }}
+      shell: bash --noprofile --norc -xeuo pipefail {0}
+      run: |
+        # Everything under this folder is packed and stored in the GitHub Cache space,
+        # and unpacked after retrieving from the cache.
+        CACHE_TMP_DIR="./cache_tmp_dir"
+        rm -rf $CACHE_TMP_DIR
+        mkdir $CACHE_TMP_DIR
+
+        # The binary archives (redist) are guaranteed to be updated as part of the release posting.
+        CTK_BASE_URL="https://developer.download.nvidia.com/compute/cuda/redist/"
+        CTK_JSON_URL="$CTK_BASE_URL/redistrib_${{ inputs.cuda-version }}.json"
+        if [[ "${{ inputs.host-platform }}" == linux* ]]; then
+          if [[ "${{ inputs.host-platform }}" == "linux-64" ]]; then
+            CTK_SUBDIR="linux-x86_64"
+          elif [[ "${{ inputs.host-platform }}" == "linux-aarch64" ]]; then
+            CTK_SUBDIR="linux-sbsa"
+          fi
+          function extract() {
+            tar -xvf $1 -C $CACHE_TMP_DIR --strip-components=1
+          }
+        elif [[ "${{ inputs.host-platform }}" == "win-64" ]]; then
+          CTK_SUBDIR="windows-x86_64"
+          function extract() {
+            _TEMP_DIR_=$(mktemp -d)
+            unzip $1 -d $_TEMP_DIR_
+            cp -r $_TEMP_DIR_/*/* $CACHE_TMP_DIR
+            rm -rf $_TEMP_DIR_
+            # see commit NVIDIA/cuda-python@69410f1d9228e775845ef6c8b4a9c7f37ffc68a5
+            chmod 644 $CACHE_TMP_DIR/LICENSE
+          }
+        fi
+        function populate_cuda_path() {
+          # take the component name as a argument
+          function download() {
+            curl -kLSs $1 -o $2
+          }
+          CTK_COMPONENT=$1
+          CTK_COMPONENT_REL_PATH="$(curl -s $CTK_JSON_URL |
+              python -c "import sys, json; print(json.load(sys.stdin)['${CTK_COMPONENT}']['${CTK_SUBDIR}']['relative_path'])")"
+          CTK_COMPONENT_URL="${CTK_BASE_URL}/${CTK_COMPONENT_REL_PATH}"
+          CTK_COMPONENT_COMPONENT_FILENAME="$(basename $CTK_COMPONENT_REL_PATH)"
+          download $CTK_COMPONENT_URL $CTK_COMPONENT_COMPONENT_FILENAME
+          extract $CTK_COMPONENT_COMPONENT_FILENAME
+          rm $CTK_COMPONENT_COMPONENT_FILENAME
+        }
+
+        # Get headers and shared libraries in place
+        for item in $(echo $CTK_CACHE_COMPONENTS | tr ',' ' '); do
+            populate_cuda_path "$item"
+        done
+        # TODO: check Windows
+        if [[ "${{ inputs.host-platform }}" == linux* ]]; then
+          mv $CACHE_TMP_DIR/lib $CACHE_TMP_DIR/lib64
+        fi
+        ls -l $CACHE_TMP_DIR
+
+        # Prepare the cache
+        # 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"
+        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
+        # Git Bash). We do it in the dumb way.
+        cp -r $CACHE_TMP_DIR/* $CUDA_PATH
+        rm -rf $CACHE_TMP_DIR
+        ls -l $CUDA_PATH
+
+    - name: Upload CTK cache
+      if: ${{ !cancelled() &&
+              steps.ctk-get-cache.outputs.cache-hit != 'true' }}
+      uses: actions/cache/save@5a3ec84eff668545956fd18022155c47e93e2684  # v4.2.3
+      with:
+        key: ${{ env.CTK_CACHE_KEY }}
+        path: ./${{ env.CTK_CACHE_FILENAME }}
+
+    - name: Restore CTK cache
+      if: ${{ steps.ctk-get-cache.outputs.cache-hit == 'true' }}
+      shell: bash --noprofile --norc -xeuo pipefail {0}
+      run: |
+        ls -l
+        CACHE_TMP_DIR="./cache_tmp_dir"
+        CUDA_PATH="./cuda_toolkit"
+        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 ${{ inputs.cuda-path }}
+
+    - name: Set output environment variables
+      shell: bash --noprofile --norc -xeuo pipefail {0}
+      run: |
+        # mimics actual CTK installation
+        if [[ "${{ inputs.host-platform }}" == linux* ]]; then
+          CUDA_PATH=$(realpath "${{ inputs.cuda-path }}")
+          echo "${CUDA_PATH}/bin" >> $GITHUB_PATH
+          echo "LD_LIBRARY_PATH=${CUDA_PATH}/lib64:${LD_LIBRARY_PATH:-}" >> $GITHUB_ENV
+        elif [[ "${{ inputs.host-platform }}" == win* ]]; then
+          function normpath() {
+            echo "$(echo $(cygpath -w $1) | sed 's/\\/\\\\/g')"
+          }
+          CUDA_PATH=$(normpath $(realpath "${{ inputs.cuda-path }}"))
+          echo "$(normpath ${CUDA_PATH}/bin)" >> $GITHUB_PATH
+        fi
+        echo "CUDA_PATH=${CUDA_PATH}" >> $GITHUB_ENV
+        echo "CUDA_HOME=${CUDA_PATH}" >> $GITHUB_ENV

.github/actions/get_pr_number/action.yml

@@ -0,0 +1,58 @@
+# SPDX-FileCopyrightText: Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+#
+# SPDX-License-Identifier: Apache-2.0
+
+name: Get the PR number
+
+description: Get the PR number without relying on the pull_request* event triggers.
+
+runs:
+  using: composite
+  steps:
+    - name: Get PR info (non-main, non-release branch)
+      if: ${{ github.ref_name != 'main' && !startsWith(github.ref_name, 'release/') }}
+      uses: nv-gha-runners/get-pr-info@main
+      id: get-pr-info
+
+    - name: Extract PR number (non-main, non-release branch)
+      if: ${{ github.ref_name != 'main' && !startsWith(github.ref_name, 'release/') }}
+      shell: bash --noprofile --norc -xeuo pipefail {0}
+      run: |
+        trap 'echo "Error at line $LINENO"; exit 1' ERR
+        PR_NUMBER="${{ fromJSON(steps.get-pr-info.outputs.pr-info).number }}"
+        if [[ -z "$PR_NUMBER" ]]; then
+          echo "Cannot extract PR number for ref: ${{ github.ref_name }}"
+          exit 1
+        fi
+        echo "PR_NUMBER=$PR_NUMBER" >> $GITHUB_ENV
+        echo "BUILD_PREVIEW=1" >> $GITHUB_ENV
+
+    - name: Get PR data (main or release/* branch)
+      if: ${{ github.ref_name == 'main' || startsWith(github.ref_name, 'release/') }}
+      uses: actions/github-script@60a0d83039c74a4aee543508d2ffcb1c3799cdea  # v7.0.1
+      id: get-pr-data
+      with:
+        script: |
+          const prs = await github.rest.repos.listPullRequestsAssociatedWithCommit({
+            commit_sha: context.sha,
+            owner: context.repo.owner,
+            repo: context.repo.repo,
+          });
+          if (!prs.data.length) {
+            core.setFailed("No PR associated with this commit on 'main' or 'release/*'.");
+          } else {
+            return prs.data[0];
+          }
+
+    - name: Extract PR number (main or release/* branch)
+      if: ${{ github.ref_name == 'main' || startsWith(github.ref_name, 'release/') }}
+      shell: bash --noprofile --norc -xeuo pipefail {0}
+      run: |
+        trap 'echo "Error at line $LINENO"; exit 1' ERR
+        PR_NUMBER="${{ fromJSON(steps.get-pr-data.outputs.result).number }}"
+        if [[ -z "$PR_NUMBER" ]]; then
+          echo "No associated PR found for the commit in 'main' or 'release/*'."
+          exit 1
+        fi
+        echo "PR_NUMBER=$PR_NUMBER" >> $GITHUB_ENV
+        echo "BUILD_LATEST=1" >> $GITHUB_ENV

.github/actions/install_unix_deps/action.yml

@@ -0,0 +1,49 @@
+# SPDX-FileCopyrightText: Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+#
+# SPDX-License-Identifier: Apache-2.0
+
+name: Install dependencies on Ubuntu
+
+description: Install needed dependencies, regardless if using GitHub- or self- hosted runners, container, sudo or not.
+
+inputs:
+  dependencies:
+    required: true
+    type: string
+  dependent_exes:
+    required: true
+    type: string
+
+runs:
+  using: composite
+  steps:
+    - name: Install dependencies
+      shell: bash --noprofile --norc -xeuo pipefail {0}
+      run: |
+        dependencies=(${{ inputs.dependencies }})
+        dependent_exes=(${{ inputs.dependent_exes }})
+
+        not_found=0
+        for dep in ${dependent_exes[@]}; do
+          if ! (command -v $dep 2>&1 >/dev/null); then
+            not_found=1
+            break
+          fi
+        done
+        if [[ $not_found == 0 ]]; then
+          echo "All dependencies are found. Do nothing."
+          exit 0
+        fi
+        if ! (command -v sudo 2>&1 >/dev/null); then
+          if [[ $EUID == 0 ]]; then
+            alias SUDO=""
+          else
+            echo "The following oprations require root access."
+            exit 1
+          fi
+        else
+          alias SUDO="sudo"
+        fi
+        shopt -s expand_aliases
+        SUDO apt update
+        SUDO apt install -y ${dependencies[@]}