CI: misc cleanup/improvement

.github/scripts/build-cpu.sh

@@ -4,12 +4,6 @@ set -xeuo pipefail
 : "${RUNNER_OS:?RUNNER_OS must be set (Linux/Windows/macOS)}"
 : "${RUNNER_ARCH:?RUNNER_ARCH must be set (X64/ARM64)}"
 
-if [[ "${RUNNER_OS}" == "Windows" ]]; then
-    pip install cmake==3.30.9
-else
-    pip install cmake==3.28.3
-fi
-
 if [ "${RUNNER_OS}" == "macOS" ] && [ "${RUNNER_ARCH}" == "ARM64" ]; then
 	cmake -DCMAKE_OSX_ARCHITECTURES=arm64 -DCOMPUTE_BACKEND=cpu .
 else

.github/scripts/build-cuda.sh

@@ -26,21 +26,18 @@ else
     [[ "${CUDA_VERSION}" == 13.*.* ]] && build_capability="75;80;86;89;90;100;120"
 fi
 
-[[ "${RUNNER_OS}" == "Windows" ]] && python3 -m pip install ninja
-
 if [ "${RUNNER_OS}" == "Linux" ]; then
     # 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 -i -w /src -v "$PWD:/src" "$image" bash -c \
         "dnf -y --refresh update --security \
-        && dnf -y install cmake gcc-toolset-11 --setopt=install_weak_deps=False --setopt=tsflags=nodocs \
+        && dnf -y install cmake gcc-toolset-11-toolchain --setopt=install_weak_deps=False --setopt=tsflags=nodocs \
         && source scl_source enable gcc-toolset-11 \
         && cmake -DCOMPUTE_BACKEND=cuda -DCOMPUTE_CAPABILITY=\"${build_capability}\" . \
-        && cmake --build . --config Release"
+        && cmake --build . --config Release --parallel"
 else
-    pip install cmake==3.28.3
     cmake -G Ninja -DCOMPUTE_BACKEND=cuda -DCOMPUTE_CAPABILITY="${build_capability}" -DCMAKE_BUILD_TYPE=Release -S .
     cmake --build . --config Release
 fi

.github/scripts/build-rocm.sh

@@ -17,10 +17,9 @@ if [ "${RUNNER_OS}" == "Linux" ]; then
     echo "Using image $image"
     docker run --rm -i \
         -w /src -v "$PWD:/src" "$image" sh -c \
-        "apt-get update \
-      && pip install cmake==3.31.6 \
+        "pip install cmake==3.31.6 \
       && cmake -DCOMPUTE_BACKEND=hip -DCMAKE_BUILD_TYPE=MinSizeRel -DCMAKE_HIP_FLAGS=\"--offload-compress\" -DBNB_ROCM_ARCH=\"${bnb_rocm_arch}\" . \
-      && cmake --build ."
+      && cmake --build . --parallel"
 else
     bnb_rocm_arch="gfx1100;gfx1101;gfx1102;gfx1150;gfx1151;gfx1200;gfx1201"
 

.github/scripts/build-xpu.sh

@@ -14,7 +14,7 @@ if [ "${RUNNER_OS}" == "Linux" ]; then
       && DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
         cmake bison intel-fw-gpu intel-ocloc \
       && cmake -DCOMPUTE_BACKEND=xpu . \
-      && cmake --build . --config Release"
+      && cmake --build . --config Release --parallel"
 fi
 
 output_dir="output/${RUNNER_OS}/X64"

.github/workflows/python-package.yml

@@ -36,7 +36,7 @@ jobs:
     steps:
       - uses: actions/checkout@v4
       - name: Setup MSVC
-        if: startsWith(matrix.os, 'windows')
+        if: runner.os == 'Windows'
         uses: ilammy/msvc-dev-cmd@v1.13.0 # to use cl
         with:
           arch: ${{ runner.arch == 'ARM64' && 'arm64' || 'x64' }}
@@ -48,6 +48,7 @@ jobs:
           name: shared_library_${{ runner.os }}_${{ runner.arch }}
           path: output/*
           retention-days: 7
+          if-no-files-found: error
 
   ##
   # This job matrix builds the CUDA versions of the libraries for platforms that support CUDA (Linux x64/aarch64 + Windows x64)
@@ -64,7 +65,7 @@ jobs:
       - uses: actions/checkout@v4
         # Windows: We install Cuda on the agent (slow)
       - uses: Jimver/cuda-toolkit@3d45d157f327c09c04b50ee6ccdea2d9d017ec76 # v0.2.35
-        if: startsWith(matrix.os, 'windows')
+        if: runner.os == 'Windows'
         id: cuda-toolkit
         with:
           cuda: ${{ matrix.cuda_version }}
@@ -75,7 +76,7 @@ jobs:
           use-local-cache: false
           log-file-suffix: ${{ runner.os }}-${{ runner.arch }}-${{matrix.cuda_version}}.txt
       - name: Setup MSVC
-        if: startsWith(matrix.os, 'windows')
+        if: runner.os == 'Windows'
         uses: ilammy/msvc-dev-cmd@v1.13.0 # to use cl
         with:
           toolset: "14.44"
@@ -89,6 +90,7 @@ jobs:
           name: shared_library_cuda_${{ runner.os }}_${{ runner.arch }}_${{ matrix.cuda_version }}
           path: output/*
           retention-days: 7
+          if-no-files-found: error
 
   build-xpu:
     strategy:
@@ -110,6 +112,7 @@ jobs:
           name: shared_library_xpu_${{ runner.os }}_${{ runner.arch }}
           path: output/*
           retention-days: 7
+          if-no-files-found: error
 
   build-rocm:
     strategy:
@@ -123,7 +126,7 @@ jobs:
     steps:
       - uses: actions/checkout@v4
       - name: Clean up disk space
-        if: startsWith(matrix.os, 'ubuntu')
+        if: runner.os == 'Linux'
         run: |
           echo "Disk space before cleanup:"
           df -h
@@ -140,7 +143,7 @@ jobs:
           echo "Disk space after cleanup:"
           df -h
       - name: Setup MSVC
-        if: startsWith(matrix.os, 'windows')
+        if: runner.os == 'Windows'
         uses: ilammy/msvc-dev-cmd@v1.13.0
         with:
           toolset: "14.44"
@@ -154,6 +157,7 @@ jobs:
           name: shared_library_rocm_${{ runner.os }}_${{ runner.arch }}_${{ matrix.rocm_version }}
           path: output/*
           retention-days: 7
+          if-no-files-found: error
 
   build-wheels:
     env:
@@ -187,11 +191,10 @@ jobs:
       - name: Set up Python
         uses: actions/setup-python@v5
         with:
-          # Python for Windows ARM64 is only available from 3.12+
-          python-version: ${{ matrix.os == 'windows-11-arm' && '3.12' || '3.10' }}
+          python-version: "3.12"
           cache: pip
       - run: pip install build wheel
-      - run: python -m build .
+      - run: python -m build -w .
       - name: Determine and Set Platform Tag, then Tag Wheel
         shell: bash
         run: |
@@ -204,6 +207,8 @@ jobs:
           name: bdist_wheel_${{ runner.os }}_${{ runner.arch }}
           path: dist/bitsandbytes-*.whl
           retention-days: 7
+          if-no-files-found: error
+          compression-level: 0
 
   upload-pre-release-wheels:
     name: Create release and upload artifacts

.github/workflows/test-runner.yml

@@ -53,82 +53,41 @@ jobs:
         id: config
         shell: bash
         run: |
-          # Map platform to OS identifiers, architecture, and test runner
+          # Map platform to test runner
           case "${{ inputs.platform }}" in
             linux-x64)
-              BUILD_OS="ubuntu-22.04"
-              ARCH="x64"
               if [[ "${{ inputs.backend }}" == "cuda" ]]; then
                 case "${{ inputs.gpu_type }}" in
-                  T4)
-                    TEST_RUNNER="bandb-aws-g4dn-4xlarge-plus-use1-public-80"
-                    ;;
-                  A10)
-                    TEST_RUNNER="bandb-aws-g5-4xlarge-plus-use1-public-80"
-                    ;;
-                  L40S)
-                    TEST_RUNNER="bandb-aws-g6e-4xlarge-plus-use1-public-80"
-                    ;;
-                  *)
-                    echo "::error::Must specify gpu_type (T4, A10, L40S) for linux-x64 cuda backend"
-                    exit 1
-                    ;;
+                  T4)    TEST_RUNNER="bandb-aws-g4dn-4xlarge-plus-use1-public-80" ;;
+                  A10)   TEST_RUNNER="bandb-aws-g5-4xlarge-plus-use1-public-80" ;;
+                  L40S)  TEST_RUNNER="bandb-aws-g6e-4xlarge-plus-use1-public-80" ;;
+                  *)     echo "::error::Must specify gpu_type (T4, A10, L40S) for linux-x64 cuda backend"; exit 1 ;;
                 esac
               else
                 case "${{ inputs.cpu_type }}" in
-                  icelake)
-                    TEST_RUNNER="banb-aws-general-8-plus-use1-public-80"
-                    ;;
-                  cascadelake)
-                    TEST_RUNNER="bandb-aws-g4dn-4xlarge-plus-use1-public-80"
-                    ;;
-                  "")
-                    TEST_RUNNER="ubuntu-22.04"
-                    ;;
-                  *)
-                    echo "::error::Invalid cpu_type: ${{ inputs.cpu_type }}"
-                    exit 1
-                    ;;
+                  icelake)     TEST_RUNNER="banb-aws-general-8-plus-use1-public-80" ;;
+                  cascadelake) TEST_RUNNER="bandb-aws-g4dn-4xlarge-plus-use1-public-80" ;;
+                  "")          TEST_RUNNER="ubuntu-22.04" ;;
+                  *)           echo "::error::Invalid cpu_type: ${{ inputs.cpu_type }}"; exit 1 ;;
                 esac
               fi
               ;;
-            linux-aarch64)
-              BUILD_OS="ubuntu-22.04-arm"
-              ARCH="aarch64"
-              TEST_RUNNER="ubuntu-22.04-arm"
-              ;;
-            macos)
-              BUILD_OS="macos-15"
-              ARCH="arm64"
-              TEST_RUNNER="macos-15"
-              ;;
+            linux-aarch64) TEST_RUNNER="ubuntu-22.04-arm" ;;
+            macos)         TEST_RUNNER="macos-15" ;;
             windows)
-              BUILD_OS="windows-2025"
-              ARCH="x64"
-              if [[ "${{ inputs.backend }}" == "cuda" ]]; then
-                TEST_RUNNER="CUDA-Windows-x64"
-              else
-                TEST_RUNNER="windows-2025"
-              fi
-              ;;
-            windows-arm64)
-              BUILD_OS="windows-11-arm"
-              ARCH="arm64"
-              TEST_RUNNER="windows-11-arm"
+              [[ "${{ inputs.backend }}" == "cuda" ]] && TEST_RUNNER="CUDA-Windows-x64" || TEST_RUNNER="windows-2025"
               ;;
+            windows-arm64) TEST_RUNNER="windows-11-arm" ;;
             *)
               echo "::error::Unsupported platform: ${{ inputs.platform }}"
               exit 1
               ;;
           esac
 
           # Create unique artifact name per configuration
-          ARTIFACT="lib_${{ inputs.backend }}_${BUILD_OS}_${ARCH}"
-          if [[ "${{ inputs.backend }}" == "cuda" ]]; then
-            ARTIFACT="${ARTIFACT}_${{ inputs.cuda_version }}_${{ inputs.gpu_type }}"
-          else
-            ARTIFACT="${ARTIFACT}_${{ inputs.cpu_type }}"
-          fi
+          ARTIFACT="lib_${{ inputs.backend }}_${RUNNER_OS}_${RUNNER_ARCH}"
+          [[ "${{ inputs.backend }}" == "cuda" ]] && ARTIFACT="${ARTIFACT}_${{ inputs.cuda_version }}_${{ inputs.gpu_type }}"
+          [[ "${{ inputs.backend }}" == "cpu"  ]] && ARTIFACT="${ARTIFACT}_${{ inputs.cpu_type }}"
           ARTIFACT="${ARTIFACT}_torch${{ inputs.torch_version }}_${{ github.run_id }}_${{ github.run_attempt }}"
 
           echo "test_runner=${TEST_RUNNER}" >> $GITHUB_OUTPUT
@@ -174,6 +133,7 @@ jobs:
           name: ${{ steps.config.outputs.artifact_name }}
           path: output/${{ runner.os }}/${{ runner.arch }}/*
           retention-days: 7
+          if-no-files-found: error
 
   test:
     needs: build