remove ref main as we may want to trigger workflows on other branches (#282)

.github/workflows/build_wheels_cuda.yml

@@ -20,8 +20,6 @@ jobs:
 
     steps:
       - uses: actions/checkout@v3
-        with:
-          ref: 'main'
 
       - uses: actions/setup-python@v3
         with:
@@ -50,7 +48,10 @@ jobs:
           $env:CUDA_PATH = $env:CONDA_PREFIX
           $env:CUDA_HOME = $env:CONDA_PREFIX
           if ($IsLinux) {$env:LD_LIBRARY_PATH = $env:CONDA_PREFIX + '/lib:' + $env:LD_LIBRARY_PATH}
+
+          # TODO: remove this
           if (!$IsLinux) {$env:INCLUDE_EXLLAMA_KERNELS = 0}
+
           $env:TORCH_CUDA_ARCH_LIST = '6.0 6.1 7.0 7.5 8.0 8.6+PTX'
           if ([decimal]$env:CUDA_VERSION -ge 11.8) { $env:TORCH_CUDA_ARCH_LIST = '6.0 6.1 7.0 7.5 8.0 8.6 8.9 9.0+PTX' }
           python setup.py sdist bdist_wheel

.github/workflows/build_wheels_pypi.yml

@@ -0,0 +1,68 @@
+name: Build AutoGPTQ Wheels for PyPI with CUDA
+
+on: workflow_dispatch
+
+jobs:
+  build_wheels:
+    if: ${{ github.repository_owner == 'PanQiWei' }}
+    name: Build wheels for ${{ matrix.os }} and Python ${{ matrix.python }} and CUDA ${{ matrix.cuda }}
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-20.04, windows-latest]
+        pyver: ["3.8", "3.9", "3.10", "3.11"]
+    defaults:
+      run:
+        shell: pwsh
+    env:
+        CUDA_VERSION: "11.7"
+
+    steps:
+      - uses: actions/checkout@v3
+
+      - uses: actions/setup-python@v3
+        with:
+          python-version: ${{ matrix.pyver }}
+
+      - name: Setup Miniconda
+        uses: conda-incubator/setup-miniconda@v2.2.0
+        with:
+          activate-environment: "build"
+          python-version: ${{ matrix.pyver }}
+          mamba-version: "*"
+          use-mamba: false
+          channels: conda-forge,defaults
+          channel-priority: true
+          add-pip-as-python-dependency: true
+          auto-activate-base: false
+
+      - name: Install Dependencies
+        run: |
+          conda install cuda-toolkit -c "nvidia/label/cuda-${env:CUDA_VERSION}.0"
+          conda install pytorch "pytorch-cuda=${env:CUDA_VERSION}" -c pytorch -c nvidia
+          python -m pip install --upgrade build setuptools wheel ninja
+
+      - name: Build Wheel
+        run: |
+          $env:CUDA_PATH = $env:CONDA_PREFIX
+          $env:CUDA_HOME = $env:CONDA_PREFIX
+          if ($IsLinux) {$env:LD_LIBRARY_PATH = $env:CONDA_PREFIX + '/lib:' + $env:LD_LIBRARY_PATH}
+
+          $env:TORCH_CUDA_ARCH_LIST = '6.0 6.1 7.0 7.5 8.0 8.6+PTX'
+          if ([decimal]$env:CUDA_VERSION -ge 11.8) { $env:TORCH_CUDA_ARCH_LIST = '6.0 6.1 7.0 7.5 8.0 8.6 8.9 9.0+PTX' }
+
+          echo "CUDA_PATH: $CUDA_PATH"
+          echo "CUDA_VERSION: $CUDA_VERSION"
+          PYPI_RELEASE=1 python setup.py sdist bdist_wheel
+
+      - uses: actions/upload-artifact@v3
+        if: runner.os == 'Linux'
+        with:
+          name: 'linux-cuda-wheels'
+          path: ./dist/*.whl
+
+      - uses: actions/upload-artifact@v3
+        if: runner.os == 'Windows'
+        with:
+          name: 'windows-cuda-wheels'
+          path: ./dist/*.whl

.github/workflows/build_wheels_rocm.yml

@@ -21,8 +21,6 @@ jobs:
 
     steps:
       - uses: actions/checkout@v3
-        with:
-          ref: 'main'
 
       - name: Free disk space
         run: |

setup.py

@@ -30,6 +30,7 @@ common_setup_kwargs = {
 }
 
 
+PYPI_RELEASE = os.environ.get('PYPI_RELEASE', None)
 BUILD_CUDA_EXT = int(os.environ.get('BUILD_CUDA_EXT', '1')) == 1
 if BUILD_CUDA_EXT:
     try:
@@ -60,8 +61,10 @@ if BUILD_CUDA_EXT:
                 "is installed without CUDA support."
             )
             sys.exit(-1)
-        common_setup_kwargs['version'] += f"+cu{CUDA_VERSION}"
 
+        # For the PyPI release, the version is simply x.x.x to comply with PEP 440.
+        if not PYPI_RELEASE:
+            common_setup_kwargs['version'] += f"+cu{CUDA_VERSION}"
 
 requirements = [
     "accelerate>=0.19.0",
@@ -110,19 +113,28 @@ if BUILD_CUDA_EXT:
         )
     ]
 
-    if os.environ.get("INCLUDE_EXLLAMA_KERNELS", "1") == "1":  # TODO: improve github action to always compile exllama_kernels
+    if os.name == "nt":
-        extensions.append(
+        # On Windows, fix an error LNK2001: unresolved external symbol cublasHgemm bug in the compilation
-            cpp_extension.CUDAExtension(
+        cuda_path = os.environ.get("CUDA_PATH", None)
-                "exllama_kernels",
+        if cuda_path is None:
-                [
+            raise ValueError("The environment variable CUDA_PATH must be set to the path to the CUDA install when installing from source on Windows systems.")
-                    "autogptq_cuda/exllama/exllama_ext.cpp",
+        extra_link_args = ["-L", f"{cuda_path}/lib/x64/cublas.lib"]
-                    "autogptq_cuda/exllama/cuda_buffers.cu",
+    else:
-                    "autogptq_cuda/exllama/cuda_func/column_remap.cu",
+        extra_link_args = []
-                    "autogptq_cuda/exllama/cuda_func/q4_matmul.cu",
+
-                    "autogptq_cuda/exllama/cuda_func/q4_matrix.cu"
+    extensions.append(
-                ]
+        cpp_extension.CUDAExtension(
-            )
+            "exllama_kernels",
+            [
+                "autogptq_cuda/exllama/exllama_ext.cpp",
+                "autogptq_cuda/exllama/cuda_buffers.cu",
+                "autogptq_cuda/exllama/cuda_func/column_remap.cu",
+                "autogptq_cuda/exllama/cuda_func/q4_matmul.cu",
+                "autogptq_cuda/exllama/cuda_func/q4_matrix.cu"
+            ],
+            extra_link_args=extra_link_args
         )
+    )
 
     additional_setup_kwargs = {
         "ext_modules": extensions,