Merge branch 'main' into build-cmake

Author: mgorny

.github/scripts/auditwheel_show.py

@@ -22,7 +22,7 @@ def main():
             print(audit_wheel_output.stdout)
 
         if audit_wheel_output.stderr:
-            print(f"**Error:**\n```{audit_wheel_output.stderr}```")
+            print(f"**Error:**\n```\n{audit_wheel_output.stderr}\n```")
 
         print("---")
 

.github/scripts/build-cpu.sh

@@ -6,12 +6,7 @@ set -xeuo pipefail
 
 pip install cmake==3.28.3
 
-if [ "${build_os:0:6}" == ubuntu ] && [ "${build_arch}" == aarch64 ]; then
-	# Allow cross-compile on aarch64
-	sudo apt-get update
-	sudo apt-get install -y gcc-aarch64-linux-gnu binutils-aarch64-linux-gnu g++-aarch64-linux-gnu
-	cmake -DCMAKE_C_COMPILER=aarch64-linux-gnu-gcc -DCMAKE_CXX_COMPILER=aarch64-linux-gnu-g++ -DCOMPUTE_BACKEND=cpu .
-elif [ "${build_os:0:5}" == macos ] && [ "${build_arch}" == aarch64 ]; then
+if [ "${build_os:0:5}" == macos ] && [ "${build_arch}" == aarch64 ]; then
 	cmake -DCMAKE_OSX_ARCHITECTURES=arm64 -DCOMPUTE_BACKEND=cpu .
 else
 	cmake -DCOMPUTE_BACKEND=cpu .

.github/scripts/build-cuda.sh

@@ -2,28 +2,38 @@
 declare build_arch
 declare build_os
 declare cuda_version
+declare cuda_targets
 
 set -xeuo pipefail
 
-# By default, target Maxwell through Hopper.
-build_capability="50;52;60;61;70;75;80;86;89;90"
+if [[ -v cuda_targets ]]; then
+    build_capability="${cuda_targets}"
+elif [ "${build_arch}" = "aarch64" ]; then
+    build_capability="75;80;90"
 
-# CUDA 11.7: Remove sm89 and sm90
-[[ "${cuda_version}" == 11.7.* ]] && build_capability="50;52;60;61;70;75;80;86"
+    # CUDA 12.8: Add sm100
+    [[ "${cuda_version}" == 12.8.* ]] && build_capability="75;80;90;100"
+else
+    # By default, target Maxwell through Hopper.
+    build_capability="50;52;60;61;70;75;80;86;89;90"
 
-# CUDA 12.8: Add sm100 and sm120; remove sm50 through sm61
-[[ "${cuda_version}" == 12.8.* ]] && build_capability="70;75;80;86;89;90;100;120"
+    # CUDA 12.8: Add sm100 and sm120; remove < sm75 to align with PyTorch 2.7+cu128 minimum
+    [[ "${cuda_version}" == 12.8.* ]] && build_capability="75;80;86;89;90;100;120"
+fi
 
 [[ "${build_os}" = windows-* ]] && python3 -m pip install ninja
 
 if [ "${build_os:0:6}" == ubuntu ]; then
-    image=nvidia/cuda:${cuda_version}-devel-ubuntu22.04
+    # We'll use Rocky Linux 8 in order to maintain manylinux 2.24 compatibility.
+    image="nvidia/cuda:${cuda_version}-devel-rockylinux8"
     echo "Using image $image"
-    docker run --platform "linux/$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 -DPTXAS_VERBOSE=1 -DCOMPUTE_BACKEND=cuda -DCOMPUTE_CAPABILITY=\"${build_capability}\" . \
-    && cmake --build ."
+
+    docker run -i -w /src -v "$PWD:/src" "$image" bash -c \
+        "dnf update -y \
+        && dnf install cmake gcc-toolset-11 -y \
+        && source scl_source enable gcc-toolset-11 \
+        && cmake -DCOMPUTE_BACKEND=cuda -DCOMPUTE_CAPABILITY=\"${build_capability}\" . \
+        && cmake --build . --config Release"
 else
     pip install cmake==3.28.3
     cmake -G Ninja -DCOMPUTE_BACKEND=cuda -DCOMPUTE_CAPABILITY="${build_capability}" -DCMAKE_BUILD_TYPE=Release -S .

.github/workflows/python-package.yml

@@ -30,14 +30,16 @@ jobs:
   build-shared-libs:
     strategy:
       matrix:
-        os: [ubuntu-22.04, macos-latest, windows-latest]
-        arch: [x86_64, aarch64]
-        exclude:
-          - os: windows-latest # This probably requires arm64 Windows agents
+        include:
+          - os: ubuntu-22.04
+            arch: x86_64
+          - os: ubuntu-22.04-arm
             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
+          - os: windows-latest
+            arch: x86_64
+          - os: macos-latest
+            arch: arm64
+    runs-on: ${{ matrix.os }}
     steps:
       - uses: actions/checkout@v4
       - name: Setup MSVC
@@ -61,24 +63,21 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        os: [ubuntu-latest, windows-latest]
-        arch: [x86_64, aarch64]
-        cuda_version:
-          ["11.7.1", "11.8.0", "12.0.1", "12.1.1", "12.2.2", "12.3.2", "12.4.1", "12.5.1", "12.6.3", "12.8.0"]
-        exclude:
-          - os: windows-latest # This probably requires arm64 Windows agents
+        os: [ubuntu-22.04, ubuntu-22.04-arm, windows-latest]
+        include:
+          - os: ubuntu-22.04
+            arch: x86_64
+          - os: ubuntu-22.04-arm
             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
+          - os: windows-latest
+            arch: x86_64
+        cuda_version:
+          ["11.8.0", "12.0.1", "12.1.1", "12.2.2", "12.3.2", "12.4.1", "12.5.1", "12.6.3", "12.8.1"]
+    runs-on: ${{ matrix.os }}
     steps:
       - uses: actions/checkout@v4
-        # Linux: We use Docker to build cross platform Cuda (aarch64 is built in emulation)
-      - name: Set up Docker multiarch
-        if: startsWith(matrix.os, 'ubuntu')
-        uses: docker/setup-qemu-action@v3
         # Windows: We install Cuda on the agent (slow)
-      - uses: Jimver/[email protected].21
+      - uses: Jimver/[email protected].22
         if: startsWith(matrix.os, 'windows')
         id: cuda-toolkit
         with:
@@ -109,17 +108,20 @@ jobs:
       - build-shared-libs-cuda
     strategy:
       matrix:
-        os: [ubuntu-latest, macos-latest, windows-latest]
+        os: [ubuntu-22.04, ubuntu-22.04-arm, windows-latest, macos-latest]
+        include:
+          - os: ubuntu-22.04
+            arch: x86_64
+          - os: ubuntu-22.04-arm
+            arch: aarch64
+          - os: windows-latest
+            arch: x86_64
+          - os: macos-latest
+            arch: arm64
         # 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
+        # code. This ensures compatibility across Python versions, including Python 3.9, 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
-            arch: aarch64
-          - os: ubuntu-latest # Temporary. Takes too long, not ready yet.
-            arch: aarch64
     runs-on: ${{ matrix.os }}
     steps:
       - uses: actions/checkout@v4
@@ -169,43 +171,137 @@ jobs:
           path: tmp/
           pattern: "bdist_wheel_*"
           merge-multiple: true
+
       - name: Inspect tmp directory after downloading artifacts
-        run: ls -alFR tmp/
+        run: |
+          ls -alFR tmp/
+          WHEEL_COUNT=$(find tmp/ -type f -name "*.whl" | wc -l)
+          echo "Found $WHEEL_COUNT wheel files"
+          if [ "$WHEEL_COUNT" -eq 0 ]; then
+            echo "::error::No wheel files found in tmp directory! Cannot proceed with release."
+            exit 1
+          fi
+
       - name: Move and rename wheel files with pattern replacement
         run: |
           mkdir -p wheels/
+
+          # The whole point of the continuous release is to have a stable download link and the only way to have a PEP 440–compliant wheel name
+          # is to use a stable placeholder version. Otherwise, pip won't let you install the wheel. The cool thing is that we can now install the
+          # wheel directly from the GH pre-release which gets updated continuously, e.g.
+          # `pip install https://github.com/bitsandbytes-foundation/bitsandbytes/releases/download/continuous-release_main/bitsandbytes-1.33.7.preview-py3-none-manylinux_2_24_x86_64.whl`
+          STABLE_PLACEHOLDER_VERSION="1.33.7.preview"
+
           # exclude macos wheels for now
           find tmp/ -type f -name '*.whl' ! -name '*macos*' -print0 | while IFS= read -r -d '' wheel; do
             wheel_filename=$(basename "$wheel")
-            # Remove the gith hash, e.g. `+1234567`, for a stable download link on the multi-backend pre-release
-            cleaned_filename=$(echo "$wheel_filename" | sed -E 's/\+[0-9a-f]{7}-/-/g')
-            mv "$wheel" "wheels/$cleaned_filename"
+
+            # Strip off the original version
+            rest=${wheel_filename#bitsandbytes-*-}
+            new_name="bitsandbytes-${STABLE_PLACEHOLDER_VERSION}-${rest}"
+
+            echo "Renaming $wheel_filename → $new_name"
+            mv "$wheel" "wheels/${new_name}"
           done
+
       - name: Inspect wheels directory after renaming files
         run: ls -alFR wheels/
-      - name: Create release and upload artifacts
-        uses: softprops/[email protected]
+
+      - uses: actions/checkout@v4
+        with:
+          path: repo
+      - name: Delete old pre-release (if exists)
+        run: |
+          cd repo && gh release delete continuous-release_main --cleanup-tag -y
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+
+      - name: Ensure tag exists
+        run: |
+          cd repo
+          git tag -f continuous-release_main
+          git push -f origin continuous-release_main
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+
+      - name: Generate pip install commands for release body
+        run: |
+          cat > body.md << 'ENDOFMARKDOWN'
+          ## Latest `main` pre-release wheel
+
+          This pre-release contains the latest development wheels for all supported platforms, rebuilt automatically on every commit to the `main` branch.
+
+          **How to install:**
+          Pick the correct command for your platform and run it in your terminal:
+
+          ENDOFMARKDOWN
+
+          for whl in wheels/*.whl; do
+            fname=$(basename "$whl")
+            url="https://github.com/bitsandbytes-foundation/bitsandbytes/releases/download/continuous-release_main/$fname"
+
+            if [[ "$fname" == *"manylinux_2_24_x86_64"* ]]; then
+              echo "### Linux (x86_64)" >> body.md
+            elif [[ "$fname" == *"manylinux_2_24_aarch64"* ]]; then
+              echo "### Linux (ARM/aarch64)" >> body.md
+            elif [[ "$fname" == *"win_amd64"* ]]; then
+              echo "### Windows (x86_64)" >> body.md
+            else
+              echo "### Other platform" >> body.md
+            fi
+
+            echo "\`\`\`sh" >> body.md
+            echo "pip install --force-reinstall $url" >> body.md
+            echo "\`\`\`" >> body.md
+            echo "" >> body.md
+          done
+
+          cat >> body.md << 'ENDOFMARKDOWN'
+          > **Note:**
+          > These wheels are updated automatically with every commit to `main` and become available as soon as the [python-package.yml](.github/workflows/python-package.yml) workflow finishes.
+
+          The version number is replaced with 1.33.7-preview in order to keep the link stable, this however does not affect the installed version at all:
+          ```
+          > pip install https://.../bitsandbytes-1.33.7-preview-py3-none-manylinux_2_24_x86_64.whl
+          Collecting bitsandbytes==1.33.7rc0
+          ...
+          Successfully installed bitsandbytes-0.46.0.dev0
+          ```
+          ENDOFMARKDOWN
+
+          # for debugging:
+          cat body.md
+
+      - name: Create new pre-release and upload artifacts
+        uses: softprops/[email protected]
         with:
           files: wheels/*.whl
           prerelease: true
           name: Latest `main` wheel
+          body_path: body.md
           tag_name: continuous-release_main
           make_latest: false
           draft: false
-          target_commitish: ${{ github.sha }}
 
   audit-wheels:
     needs: build-wheels
-    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        os: [ubuntu-22.04, ubuntu-22.04-arm]
+        include:
+          - os: ubuntu-22.04
+            arch: x86_64
+          - os: ubuntu-22.04-arm
+            arch: aarch64
+    runs-on: ${{ matrix.os }}
     env:
       PIP_DISABLE_PIP_VERSION_CHECK: 1
     steps:
       - uses: actions/checkout@v4
-      - name: Download all wheels
+      - name: Download wheel
         uses: actions/download-artifact@v4
         with:
-          merge-multiple: true
-          pattern: "bdist_wheel_*"
+          name: bdist_wheel_${{ matrix.os }}_${{ matrix.arch }}
           path: wheels/
       - name: Set up Python
         uses: actions/setup-python@v5
@@ -241,33 +337,3 @@ jobs:
         uses: pypa/gh-action-pypi-publish@release/v1
         with:
           print-hash: true
-
-#  test:
-#    needs:
-#      - build-wheels
-#    strategy:
-#      fail-fast: false
-#      matrix:
-#        include:
-#          - os: ubuntu-latest
-#            arch: x86_64
-#            python-version: "3.8"
-#          - os: windows-latest
-#            arch: x86_64
-#            python-version: "3.8"
-#    runs-on: ${{ matrix.os }}
-#    steps:
-#      - uses: actions/checkout@v4
-#      - uses: actions/download-artifact@v4
-#        with:
-#          merge-multiple: true
-#          pattern: "bdist_wheel_${{ matrix.os }}_${{ matrix.arch }}*"
-#          path: wheel/
-#      - uses: actions/setup-python@v5
-#        with:
-#          python-version: ${{ matrix.python-version }}
-#          cache: pip
-#      - shell: bash
-#        run: ls -lar wheel/
-#      - run: pip install wheel/*.whl -r requirements-ci.txt
-#      - run: pytest --log-cli-level=DEBUG --continue-on-collection-errors tests