[ci]: Refactor CI workflows to use Miniconda for CUDA setup, enhancing environment management and installation verification

Author: haoyu-zc

.github/workflows/build_wheels.yml

@@ -22,31 +22,38 @@ jobs:
     steps:
       - uses: actions/checkout@v4
 
-      # Install CUDA on the runner
-      - name: Install CUDA
-        uses: Jimver/cuda-toolkit@v0.2.16
-        id: cuda-toolkit
+      # Set up conda with CUDA environment
+      - name: Setup Miniconda
+        uses: conda-incubator/setup-miniconda@v3
         with:
-          cuda: ${{ matrix.cuda-version }}
-          method: 'network'
-          sub-packages: '["nvcc", "cudart", "cufft", "cublas", "cusparse", "cusolver", "curand", "npp", "nvrtc"]'
-
-      - name: Set up Python
-        uses: actions/setup-python@v5
-        with:
-          python-version: '3.11'  # Version for running cibuildwheel
+          environment-file: environment/environment-gpu.yml
+          activate-environment: ccc-gpu
+          auto-activate-base: false
+          miniforge-variant: Mambaforge
+          miniforge-version: latest
+          use-mamba: true
+
+      - name: Verify CUDA installation
+        shell: bash -l {0}
+        run: |
+          conda info
+          conda list cuda
+          nvcc --version
+          echo "CUDA_HOME: $CONDA_PREFIX"
 
       - name: Install cibuildwheel
+        shell: bash -l {0}
         run: python -m pip install cibuildwheel==2.21.3
 
       - name: Build wheels
+        shell: bash -l {0}
         env:
           CIBW_BUILD: "cp310-manylinux_x86_64 cp311-manylinux_x86_64 cp312-manylinux_x86_64 cp313-manylinux_x86_64 cp314-manylinux_x86_64"
           # Skip PyPy, musllinux, and other architectures
           CIBW_SKIP: "pp* *-musllinux* *-manylinux_i686 *-manylinux_aarch64"
           # Install CUDA and other dependencies inside the container
           CIBW_BEFORE_ALL_LINUX: |
-            # Install CUDA repository and toolkit
+            # Install CUDA repository and toolkit - using same version as conda env
             yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/cuda-rhel7.repo
             yum clean all
             yum install -y cuda-toolkit-12-5 ninja-build
@@ -118,16 +125,16 @@ jobs:
     steps:
       - uses: actions/checkout@v4
 
-      # Install CUDA for testing
-      - name: Install CUDA
-        uses: Jimver/cuda-toolkit@v0.2.16
-        with:
-          cuda: ${{ matrix.cuda-version }}
-          method: 'network'
-
-      - name: Set up Python ${{ matrix.python-version }}
-        uses: actions/setup-python@v5
+      # Set up conda with CUDA environment for testing
+      - name: Setup Miniconda
+        uses: conda-incubator/setup-miniconda@v3
         with:
+          environment-file: environment/environment-gpu.yml
+          activate-environment: ccc-gpu
+          auto-activate-base: false
+          miniforge-variant: Mambaforge
+          miniforge-version: latest
+          use-mamba: true
           python-version: ${{ matrix.python-version }}
 
       - name: Download wheels
@@ -137,17 +144,19 @@ jobs:
           path: ./wheelhouse
 
       - name: Install wheel and test dependencies
+        shell: bash -l {0}
         run: |
           python -m pip install --upgrade pip
-          python -m pip install numpy scipy numba pandas scikit-learn pytest
           python -m pip install --find-links ./wheelhouse cccgpu
 
       - name: Run basic import test
+        shell: bash -l {0}
         run: |
           python -c "import ccc; import ccc_cuda_ext; print('All imports successful')"
           python -c "from ccc import compute; print('CCC compute module loaded')"
 
       - name: Run pytest
+        shell: bash -l {0}
         run: |
           pytest tests/test_coef.py -v
 

.github/workflows/build_wheels_simple.yml

@@ -23,22 +23,28 @@ jobs:
     steps:
       - uses: actions/checkout@v4
 
-      - name: Install CUDA
-        uses: Jimver/cuda-toolkit@v0.2.16
-        id: cuda-toolkit
+      - name: Setup Miniconda
+        uses: conda-incubator/setup-miniconda@v3
         with:
-          cuda: '12.5.0'
-          method: 'network'
+          environment-file: environment/environment-gpu.yml
+          activate-environment: ccc-gpu
+          auto-activate-base: false
+          miniforge-variant: Mambaforge
+          miniforge-version: latest
+          use-mamba: true
 
-      - name: Set up Python
-        uses: actions/setup-python@v5
-        with:
-          python-version: '3.11'
+      - name: Verify CUDA installation
+        shell: bash -l {0}
+        run: |
+          conda info
+          nvcc --version
 
       - name: Install cibuildwheel
+        shell: bash -l {0}
         run: python -m pip install cibuildwheel==2.21.3
 
       - name: Build wheel
+        shell: bash -l {0}
         env:
           CIBW_BUILD: "${{ inputs.python_version }}-*"
           CIBW_SKIP: "pp* *-musllinux* *-manylinux_i686 *-manylinux_aarch64 *-win32"

docs/WHEEL_BUILDING.md

@@ -18,12 +18,13 @@ This is the primary workflow that builds wheels for all supported configurations
 - **CUDA version**: 12.5 (for stable GPU support and features)
 
 #### Key Features:
-1. Builds wheels for all Python versions in parallel on Linux
-2. Automatically installs CUDA toolkit in manylinux container
-3. Bundles CUDA libraries with wheels for standalone distribution
-4. Runs basic import tests after building
-5. Uploads wheels as GitHub Actions artifacts
-6. Optionally publishes to PyPI on version tags
+1. Uses conda to manage CUDA 12.5 environment for reliable installation
+2. Builds wheels for all Python versions in parallel on Linux
+3. Automatically installs CUDA toolkit in manylinux container
+4. Bundles CUDA libraries with wheels for standalone distribution
+5. Runs basic import tests after building
+6. Uploads wheels as GitHub Actions artifacts
+7. Optionally publishes to PyPI on version tags
 
 ### Simplified Workflow: `build_wheels_simple.yml`
 
@@ -54,12 +55,23 @@ Helper script for Windows wheels:
 
 ### Prerequisites
 
-1. Install cibuildwheel:
+1. Install conda/mamba:
 ```bash
-pip install cibuildwheel
+# Install miniforge (includes mamba)
+curl -L -O "https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-$(uname)-$(uname -m).sh"
+bash Miniforge3-$(uname)-$(uname -m).sh
+```
+
+2. Create CUDA environment:
+```bash
+mamba env create -f environment/environment-gpu.yml
+conda activate ccc-gpu
 ```
 
-2. Install CUDA Toolkit 12.5 (or compatible version)
+3. Install cibuildwheel:
+```bash
+pip install cibuildwheel
+```
 
 ### Build Commands
 
@@ -73,21 +85,25 @@ Build specific Python version:
 CIBW_BUILD="cp311-*" cibuildwheel --output-dir wheelhouse
 ```
 
-Build with custom CUDA path (Linux):
+Build with conda CUDA environment:
 ```bash
-export CUDA_HOME=/usr/local/cuda-12.5
+# Activate the conda environment with CUDA
+conda activate ccc-gpu
+# CUDA paths are automatically set by conda
 cibuildwheel --output-dir wheelhouse
 ```
 
 ## Platform-Specific Notes
 
 ### Linux (manylinux)
 
-- Uses manylinux2014 base image for broad Linux distribution compatibility
-- CUDA 12.5 is installed inside the container via yum package manager
+- Uses conda to install CUDA 12.5 on the host system for reliable setup
+- Uses manylinux2014 base image for broad Linux distribution compatibility  
+- CUDA 12.5 is also installed inside the container via yum package manager
 - Wheels are repaired with `auditwheel` to bundle CUDA shared libraries
 - Resulting wheels are compatible with most Linux distributions (CentOS 7+, Ubuntu 16.04+, etc.)
 - Supports x86_64 architecture only
+- Conda environment ensures consistent CUDA version across local and CI builds
 
 ## CUDA Architecture Support
 
@@ -119,8 +135,9 @@ To support additional architectures, modify the `CMAKE_CUDA_ARCHITECTURES` setti
 ### Common Issues
 
 1. **CUDA not found during build**
-   - Ensure CUDA_HOME/CUDA_PATH environment variable is set
-   - Verify CUDA toolkit version compatibility
+   - Ensure conda environment is activated: `conda activate ccc-gpu`
+   - Verify CUDA installation: `nvcc --version`
+   - Check conda CUDA package: `conda list cuda`
 
 2. **Wheel import fails with missing shared library**
    - Check auditwheel output for excluded libraries
@@ -134,6 +151,9 @@ To support additional architectures, modify the `CMAKE_CUDA_ARCHITECTURES` setti
 
 Enable verbose output:
 ```bash
+# Ensure conda environment is active
+conda activate ccc-gpu
+# Enable verbose cibuildwheel output
 CIBW_BUILD_VERBOSITY=3 cibuildwheel --output-dir wheelhouse
 ```