name: Unit tests

on:
  workflow_dispatch:
  schedule:
    # Every day at 02:15 AM UTC
    - cron: "15 2 * * *"

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

jobs:

  build-cpu:
    strategy:
      matrix:
        os: [ubuntu-22.04, ubuntu-22.04-arm, windows-2025, macos-15]
        include:
          - os: ubuntu-22.04
            arch: x86_64
          - os: ubuntu-22.04-arm
            arch: aarch64
          - os: windows-2025
            arch: x86_64
          - os: macos-15
            arch: arm64
    runs-on: ${{ matrix.os }}
    steps:
      - uses: actions/checkout@v4

      - name: Setup MSVC
        if: startsWith(matrix.os, 'windows')
        uses: ilammy/msvc-dev-cmd@v1.13.0 # to use cl

      - name: Build C++
        run: bash .github/scripts/build-cpu.sh
        env:
          build_os: ${{ matrix.os }}
          build_arch: ${{ matrix.arch }}

      - name: Upload build artifact
        uses: actions/upload-artifact@v4
        with:
          name: lib_cpu_${{ matrix.os }}_${{ matrix.arch }}
          path: output/${{ matrix.os }}/${{ matrix.arch }}/*
          retention-days: 7

  build-cuda:
    strategy:
      matrix:
        cuda_version: ["11.8.0", "12.6.3", "12.8.1"]
        os: [ubuntu-22.04, ubuntu-22.04-arm, windows-2025]
        include:
          - os: ubuntu-22.04
            arch: x86_64
          - os: ubuntu-22.04-arm
            arch: aarch64
          - os: windows-2025
            arch: x86_64
    runs-on: ${{ matrix.os }}

    steps:
      - uses: actions/checkout@v4

      - name: Install CUDA Toolkit
        uses: Jimver/cuda-toolkit@v0.2.23
        if: startsWith(matrix.os, 'windows')
        id: cuda-toolkit
        with:
          cuda: ${{ matrix.cuda_version }}
          method: "network"
          sub-packages: '["nvcc","cudart","cusparse","cublas","thrust","nvrtc_dev","cublas_dev","cusparse_dev"]'
          use-github-cache: false

      - name: Setup MSVC
        if: startsWith(matrix.os, 'windows')
        uses: ilammy/msvc-dev-cmd@v1.13.0 # to use cl

      # Target just the architectures used for our runners.
      - name: Build C++ / CUDA
        run: bash .github/scripts/build-cuda.sh
        env:
          build_os: ${{ matrix.os }}
          build_arch: ${{ matrix.arch }}
          cuda_version: ${{ matrix.cuda_version }}
          cuda_targets: "75;89"

      - name: Upload build artifact
        uses: actions/upload-artifact@v4
        with:
          name: lib_cuda_${{matrix.cuda_version}}_${{ matrix.os }}_${{ matrix.arch }}
          path: output/${{ matrix.os }}/${{ matrix.arch }}/*
          retention-days: 7

  test-cpu:
    if: github.repository == 'bitsandbytes-foundation/bitsandbytes'
    needs: build-cpu
    strategy:
      fail-fast: false
      matrix:
        os: [ubuntu-22.04, ubuntu-22.04-arm, windows-2025, macos-15]
        # Test with the oldest supported torch version and the two newest.
        torch_version: ["2.2.2", "2.6.0", "2.7.1"]
        include:
          - os: ubuntu-22.04
            arch: x86_64
            runner: banb-aws-general-8-plus-use1-public-80
          - os: ubuntu-22.04-arm
            arch: aarch64
          - os: ubuntu-22.04-arm
            arch: aarch64
            torch_version: "2.5.1"
          - os: windows-2025
            arch: x86_64
          - os: macos-15
            arch: arm64
        exclude:
          - os: ubuntu-22.04-arm
            torch_version: "2.2.2"

    runs-on: ${{ matrix.runner || matrix.os }}
    env:
      BNB_TEST_DEVICE: cpu
    steps:
      - uses: actions/checkout@v4

      - name: Download build artifact
        uses: actions/download-artifact@v4
        with:
          name: lib_cpu_${{ matrix.os }}_${{ matrix.arch }}
          path: bitsandbytes/
          merge-multiple: true

      - name: Setup Python
        uses: actions/setup-python@v5
        with:
          python-version: 3.9

      - name: Setup MSVC
        if: startsWith(matrix.os, 'windows')
        uses: ilammy/msvc-dev-cmd@v1.13.0 # to use cl for torch.compile

      - name: Install dependencies
        run: |
          pip install torch==${{ matrix.torch_version }} --index-url https://download.pytorch.org/whl/cpu
          pip install -e ".[test]"
          pip install pytest-cov

      # We need to downgrade to numpy<2 for torch<2.3 compatibility.
      - name: Downgrade NumPy
        if: startsWith(matrix.torch_version, '2.2.')
        run: pip install "numpy<2"

      - name: Show installed packages
        run: pip list

      - name: Show environment information
        run: python -m torch.utils.collect_env

      - name: Run tests
        run: pytest --durations=100

  test-cpu-ipex:
    if: github.repository == 'bitsandbytes-foundation/bitsandbytes'
    needs: build-cpu
    runs-on: banb-aws-general-8-plus-use1-public-80
    env:
      BNB_TEST_DEVICE: cpu
    steps:
      - uses: actions/checkout@v4

      - name: Download build artifact
        uses: actions/download-artifact@v4
        with:
          name: lib_cpu_ubuntu-22.04_x86_64
          path: bitsandbytes/
          merge-multiple: true

      - name: Setup Python
        uses: actions/setup-python@v5
        with:
          python-version: 3.9

      - name: Install dependencies
        run: |
          pip install torch==2.7.1 --index-url https://download.pytorch.org/whl/cpu
          pip install intel_extension_for_pytorch==2.7.0 --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/cpu/us/
          pip install -e ".[test]"
          pip install pytest-cov

      - name: Show installed packages
        run: pip list

      - name: Show environment information
        run: python -m torch.utils.collect_env

      - name: IPEX smoke test
        run: python -c "import torch; import intel_extension_for_pytorch as ipex; print(torch.__version__); print(ipex.__version__);"

      - name: Run tests
        run: pytest --durations=100

  # test-cuda-aarch64:
  #   if: github.repository == 'bitsandbytes-foundation/bitsandbytes'
  #   needs: build-cuda
  #   strategy:
  #     fail-fast: false
  #     matrix:
  #       os: [ubuntu-22.04-arm]
  #       arch: [aarch64]
  #       torch_version: ["2.7.0"]
  #       cuda_version: ["11.8.0", "12.8.1"]

  #   runs-on: bandb-aws-g5g-4xlarge-plus-use1-public-80
  #   env:
  #     BNB_TEST_DEVICE: cuda
  #   steps:
  #     - name: Show GPU Information
  #       run: nvidia-smi

  #     - name: Show pip packages
  #       run: pip list

  test-cuda:
    if: github.repository == 'bitsandbytes-foundation/bitsandbytes'
    needs: build-cuda
    strategy:
      fail-fast: false
      matrix:
        os: [ubuntu-22.04, windows-2025]
        arch: [x86_64]
        gpu: [T4, L40S]
        cuda_version: ["11.8.0", "12.6.3", "12.8.1"]
        include:
          - cuda_version: "11.8.0"
            torch_version: "2.2.2"
            pypi_index: "https://download.pytorch.org/whl/cu118"
          - cuda_version: "12.6.3"
            torch_version: "2.6.0"
            pypi_index: "https://download.pytorch.org/whl/cu126"
          - cuda_version: "12.8.1"
            torch_version: "2.7.1"
            pypi_index: "https://download.pytorch.org/whl/cu128"


          # Linux L40S runners
          - os: ubuntu-22.04
            gpu: L40S
            runner: bandb-aws-g6e-4xlarge-plus-use1-public-80

          # Linux T4 runners
          - os: ubuntu-22.04
            gpu: T4
            runner: bandb-aws-g4dn-4xlarge-plus-use1-public-80

          # Specific Windows runners using cu118
          - os: windows-2025
            arch: x86_64
            gpu: T4
            runner: CUDA-Windows-x64
            cuda_version: "11.8.0"
            torch_version: "2.2.0"
            pypi_index: "https://download.pytorch.org/whl/cu118"
          - os: windows-2025
            arch: x86_64
            gpu: T4
            runner: CUDA-Windows-x64
            cuda_version: "11.8.0"
            torch_version: "2.6.0"
            pypi_index: "https://download.pytorch.org/whl/cu118"
          - os: windows-2025
            arch: x86_64
            gpu: T4
            runner: CUDA-Windows-x64
            cuda_version: "11.8.0"
            torch_version: "2.7.1"
            pypi_index: "https://download.pytorch.org/whl/cu118"

        exclude:
          # Our current T4 Windows runner has a driver too old (471.11)
          # and cannot support CUDA 12+. Skip for now.
          - os: windows-2025
            cuda_version: "12.8.1"
          - os: windows-2025
            cuda_version: "12.6.3"

          # No Windows L40S runners.
          - os: windows-2025
            gpu: L40S
    runs-on: ${{ matrix.runner }}
    env:
      BNB_TEST_DEVICE: cuda
    steps:
      - name: Show GPU Information
        run: nvidia-smi

      - uses: actions/checkout@v4

      - name: Download build artifact
        uses: actions/download-artifact@v4
        with:
          name: lib_cuda_${{ matrix.cuda_version }}_${{ matrix.os }}_${{ matrix.arch }}
          path: bitsandbytes/
          merge-multiple: true

      - name: Setup Python
        uses: actions/setup-python@v5
        with:
          python-version: 3.9

      - name: Install dependencies
        run: |
          pip install torch==${{ matrix.torch_version }} --index-url ${{ matrix.pypi_index }}
          pip install -e ".[test]"
          pip install pytest-cov

        # We need to downgrade to numpy<2 for torch<2.3 compatibility.
      - name: Downgrade NumPy
        if: startsWith(matrix.torch_version, '2.2.')
        run: pip install "numpy<2"

      - name: Show installed packages
        run: pip list

      - name: Show environment information
        run: python -m torch.utils.collect_env

      - name: Run tests
        run: pytest --durations=100