Build: Expand CUDA Toolkit Matrix (#1111)

* (ci) build with wider CUDA version matrix

* (ci) build with wider CUDA version matrix

* (ci) skip sm_89 target on CUDA 11.7

* (ci) skip sm_90 target on CUDA 11.8

* modify workflow to publish to test.pypi

* (build) Test for manylinux_2_24 build on GH actions

* (build) got that backwards.

* try fixing manual triggering condition for testpypi

* try if Ubuntu 18.04 is an easy fix to allow for `manylinux_2_24` compatibility

* hardcode publish step to run to test publishing

* set ubuntu to newest supported version

* try statically linking libstdc++ to achieve manylinux_2_18

* last commit only brought us to manylinux_2_34, reverse

* add misssing permission for publishing to pypi

* snake case deprecated in favor of kebab

* downgrade cuda ubuntu aiming for manylinux_2_24

* add step to upgrade cmake due to old Ubuntu for CUDA build

* adjust path to prefer pip installed cmake

* (cmake) set CMAKE_BUILD_TYPE=Release if unspecified

* default to CMAKE_BUILD_TYPE Release for optimized releases and better many_linux compatibility

* (build) back to ubuntu22.04 docker images

* verify Cmake in separte step

* add clarifying comment about Python version compatibility

* (build) we don't need cmake for wheel step

* fixup testpypi publish to run in PR for testing

* add pypi publishing when tagged on main

* add functionality to rewrite platform tags

* (ci) adjust platform tags for wheels

* fix for windows, get order right.

* fix for windows, get order right.

* (build) slim down those fatbins on windows cuda

* sloppy

* remove broken PyPi upload for now

---------
Co-authored-by: Titus von Koeller <9048635+Titus-von-Koeller@users.noreply.github.com>

.github/workflows/python-package.yml

@@ -17,6 +17,7 @@ on:
       - 'pytest.ini'
   release:
     types: [ published ]
+  workflow_dispatch: {}  # Allow manual trigger
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}
@@ -82,10 +83,12 @@ jobs:
       matrix:
         os: [ubuntu-latest, windows-latest]
         arch: [x86_64, aarch64]
-        cuda_version: ['12.1.0']
+        cuda_version: ["11.7.1", "11.8.0", "12.0.1", "12.1.1", "12.2.2", "12.3.2"]
         exclude:
           - os: windows-latest # This probably requires arm64 Windows agents
             arch: aarch64
+          - os: ubuntu-latest # Temporary. Takes too long, not ready yet.
+            arch: aarch64
     runs-on: ${{ matrix.os }} # One day, we could run them on native agents. Azure supports this now but it's planned only for Q3 2023 for hosted agents
     steps:
       # Check out code
@@ -121,6 +124,9 @@ jobs:
         set -ex
         build_os=${{ matrix.os }}
         build_arch=${{ matrix.arch }}
+        build_capability="50;52;60;61;70;75;80;86;89;90"
+        [[ "${{ matrix.cuda_version }}" == 11.7.* ]] && build_capability=${build_capability%??????}
+        [[ "${{ matrix.cuda_version }}" == 11.8.* ]] && build_capability=${build_capability%???}
         [[ "${{ matrix.os }}" = windows-* ]] && python3 -m pip install ninja
         for NO_CUBLASLT in ON OFF; do
           if [ ${build_os:0:6} == ubuntu ]; then
@@ -129,10 +135,10 @@ jobs:
             docker run --platform linux/$build_arch -i -w /src -v $PWD:/src $image sh -c \
               "apt-get update \
               && DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends cmake \
-              && cmake -DCOMPUTE_BACKEND=cuda -DCOMPUTE_CAPABILITY=\"50;52;60;61;70;75;80;86;89;90\" -DNO_CUBLASLT=${NO_CUBLASLT} . \
+              && cmake -DCOMPUTE_BACKEND=cuda -DCOMPUTE_CAPABILITY=\"${build_capability}\" -DNO_CUBLASLT=${NO_CUBLASLT} . \
               && cmake --build ."
           else
-            cmake -G Ninja -DCOMPUTE_BACKEND=cuda -DNO_CUBLASLT=${NO_CUBLASLT} -DCMAKE_BUILD_TYPE=Release -S .
+            cmake -G Ninja -DCOMPUTE_BACKEND=cuda -DCOMPUTE_CAPABILITY="${build_capability}" -DNO_CUBLASLT=${NO_CUBLASLT} -DCMAKE_BUILD_TYPE=Release -S .
             cmake --build . --config Release
           fi
         done
@@ -151,7 +157,10 @@ jobs:
     strategy:
       matrix:
         os: [ubuntu-latest, macos-latest, windows-latest]
-        python-version: ["3.9", "3.10", "3.11", "3.12"]
+        # The specific Python version is irrelevant in this context as we are only packaging non-C extension
+        # code. This ensures compatibility across Python versions, including Python 3.8, as compatibility is
+        # dictated by the packaged code itself, not the Python version used for packaging.
+        python-version: ["3.10"]
         arch: [x86_64, aarch64]
         exclude:
           - os: windows-latest # This probably requires arm64 Windows agents
@@ -192,27 +201,15 @@ jobs:
     - name: Build wheel
       shell: bash
       run: python -m build .
+    - name: Determine and Set Platform Tag, then Tag Wheel
+      shell: bash
+      run: |
+        PLATFORM_TAG=$(python scripts/set_platform_tag.py ${{ matrix.arch }})
+        echo "PLATFORM_TAG=$PLATFORM_TAG"
+        wheel tags --remove --abi-tag=none --python-tag=py3 --platform-tag=$PLATFORM_TAG dist/bitsandbytes-*.whl
     - name: Upload build artifact
       uses: actions/upload-artifact@v4
       with:
-        name: bdist_wheel_${{ matrix.os }}_${{ matrix.arch }}_${{ matrix.python-version }}
+        name: bdist_wheel_${{ matrix.os }}_${{ matrix.arch }}
         path: dist/bitsandbytes-*.whl
         retention-days: 7
-  publish:
-    needs: build-wheels
-    runs-on: ubuntu-latest
-    steps:
-    - uses: actions/checkout@v4
-    - name: Download build artifact
-      uses: actions/download-artifact@v4
-      with:
-        path: dist/
-        merge-multiple: true
-        pattern: "bdist_wheel_*"
-    - run: |
-        ls -lR dist/
-    - name: Publish to PyPi
-      if: startsWith(github.ref, 'refs/tags')
-      uses: pypa/gh-action-pypi-publish@release/v1
-      with:
-        password: ${{ secrets.pypi }}

CMakeLists.txt

@@ -15,6 +15,14 @@ cmake_minimum_required(VERSION 3.22.1)
 
 project(bitsandbytes LANGUAGES CXX)
 
+# If run without specifying a build type, default to using the Release configuration:
+#    optimizing the generated binaries for performance and also adds the `-DNDEBUG` flag,
+#    which turns off a bunch of asserts which seem to link to new symbols in libstdc++,
+#    worsening our many_linux compliance..
+if(NOT CMAKE_BUILD_TYPE)
+    set(CMAKE_BUILD_TYPE Release)
+endif()
+
 # Define included source files
 set(CPP_FILES csrc/common.cpp csrc/cpu_ops.cpp csrc/pythonInterface.cpp)
 set(CUDA_FILES csrc/ops.cu csrc/kernels.cu)
@@ -108,6 +116,7 @@ if(BUILD_CUDA)
     endif()
 
     string(APPEND CMAKE_CUDA_FLAGS " --use_fast_math")
+
     if(PTXAS_VERBOSE)
         # Verbose? Outputs register usage information, and other things...
         string(APPEND CMAKE_CUDA_FLAGS " -Xptxas=-v")
@@ -220,4 +229,4 @@ if(MSVC)
     set_target_properties(bitsandbytes PROPERTIES RUNTIME_OUTPUT_DIRECTORY_DEBUG "${PROJECT_SOURCE_DIR}/bitsandbytes")
 endif()
 
-set_target_properties(bitsandbytes PROPERTIES LIBRARY_OUTPUT_DIRECTORY bitsandbytes)
+set_target_properties(bitsandbytes PROPERTIES LIBRARY_OUTPUT_DIRECTORY "${PROJECT_SOURCE_DIR}/bitsandbytes")

scripts/set_platform_tag.py

+import argparse
+import platform
+import sys
+
+
+def get_platform_tag(architecture):
+    system = platform.system()
+
+    if system == "Linux":
+        tag = (
+            "manylinux_2_24_x86_64" if architecture == "x86_64" else "manylinux_2_24_aarch64"
+        )
+    elif system == "Darwin":
+        tag = "macosx_13_1_x86_64" if architecture == "x86_64" else "macosx_13_1_arm64"
+    elif system == "Windows":
+        tag = "win_amd64" if architecture == "x86_64" else "win_arm64"
+    else:
+        sys.exit(f"Unsupported system: {system}")
+
+    return tag
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Determine platform tag.")
+    parser.add_argument("arch", type=str, help="Architecture (e.g., x86_64, aarch64)")
+    args = parser.parse_args()
+
+    tag = get_platform_tag(args.arch)
+
+    print(tag)  # This will be captured by the GitHub Actions workflow
+
+
+if __name__ == "__main__":
+    main()