ci-gpu_tests.yml

name: GPU Tests

# Builds Cytnx with CUDA (cuTENSOR + cuQuantum) and runs the gpu_test_main
# gtest suite on the self-hosted GPU runner. This is a *native* build: the
# runner already provides the CUDA toolkit, cuTENSOR, cuQuantum, OpenBLAS and
# gtest system-wide, so unlike a GitHub-hosted runner there is no conda/PyPI
# toolchain install step.
#
# Runs on manual dispatch, on same-repo pull requests, and on push to master.
# The push trigger covers code merged from a fork PR (which the pull_request
# run deliberately skips), exercising the trusted merged revision on master.
# Fork pull requests are NOT executed on the self-hosted runner: on a public
# repo that would let a fork run arbitrary code on our hardware. The job-level
# `if` enforces this (a push event is never a fork, so it always runs).
on:
  workflow_dispatch:
  push:
    branches:
      - master
  pull_request:
    branches:
      - master

concurrency:
  group: gpu-tests-${{ github.ref }}
  cancel-in-progress: true

jobs:
  build-and-test:
    # Manual runs always; PRs only when the head branch is in this repo (not a
    # fork). Never let fork PR code run on the self-hosted GPU runner.
    if: >-
      github.event_name != 'pull_request' ||
      github.event.pull_request.head.repo.full_name == github.repository
    # Matches the self-hosted runner's labels (see `gh api .../actions/runners`).
    runs-on: [self-hosted, linux, x64, cuda, gpu]
    timeout-minutes: 120
    defaults:
      run:
        shell: bash -eo pipefail {0}
    env:
      # cuTENSOR / cuQuantum install roots on the runner. Kept as repository
      # Actions variables so this file carries no machine-specific paths; the
      # FindCUTENSOR / FindCUQUANTUM modules require both to be set.
      CUTENSOR_ROOT: ${{ vars.CUTENSOR_ROOT }}
      CUQUANTUM_ROOT: ${{ vars.CUQUANTUM_ROOT }}
      # The runner's GPU is an RTX 4070 Ti SUPER (Ada, sm_89). CUDA 13 dropped
      # sm_70, so the repo-default fatbin can't be used as-is; pin per runner.
      GPU_CUDA_ARCH: ${{ vars.GPU_CUDA_ARCH || '89' }}
      BUILD_DIR: ${{ github.workspace }}/build_gpu
    steps:
      - uses: actions/checkout@v4
        with:
          fetch-depth: 0
          submodules: recursive

      # When re-running a PR job, GitHub Actions reuses the merge commit from the
      # original run. Re-merge the current base so we always test against its tip.
      - name: Merge with latest target branch (pull_request only)
        if: github.event_name == 'pull_request'
        run: |
          git config user.name "github-actions[bot]"
          git config user.email "github-actions[bot]@users.noreply.github.com"
          git fetch origin ${{ github.event.pull_request.base.ref }}
          git merge --no-edit origin/${{ github.event.pull_request.base.ref }}
          git submodule update --init --recursive

      - name: Put CUDA toolkit on PATH
        run: echo "/usr/local/cuda/bin" >> "$GITHUB_PATH"

      - name: GPU and toolkit info
        run: |
          nvidia-smi
          nvcc --version

      - name: Configure (CUDA Release, C++ tests, no Python)
        run: >-
          cmake -S . -B "$BUILD_DIR"
          -DCMAKE_BUILD_TYPE=Release
          -DUSE_CUDA=ON -DUSE_CUTENSOR=ON -DUSE_CUQUANTUM=ON
          -DRUN_TESTS=ON -DBUILD_PYTHON=OFF
          -DCMAKE_CUDA_ARCHITECTURES="$GPU_CUDA_ARCH"

      - name: Build gpu_test_main
        run: cmake --build "$BUILD_DIR" -j "$(nproc)" --target gpu_test_main

      # `-L gpu` selects only the tests tagged with the `gpu` label added in
      # tests/gpu/CMakeLists.txt (the CPU suite is already covered by
      # ci-cmake_tests.yml, so there is no need to re-run it here).
      - name: Run GPU test suite
        run: GTEST_COLOR=1 ctest --test-dir "$BUILD_DIR" -L gpu --output-on-failure