From be785d760b63ba60adaa40171f7707ccd9b8fff0 Mon Sep 17 00:00:00 2001
From: Matthew Douglas <38992547+matthewdouglas@users.noreply.github.com>
Date: Mon, 15 Jun 2026 12:59:04 -0400
Subject: [PATCH 01/11] Try improving VS2026 + CUDA < 13.2 compat.

---
 CMakeLists.txt | 20 +++++++++++++++++---
 1 file changed, 17 insertions(+), 3 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 3d420edb1..ccf0a5938 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -55,6 +55,10 @@ if(NOT CMAKE_BUILD_TYPE)
     set(CMAKE_BUILD_TYPE Release)
 endif()
 
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
+set(CMAKE_CXX_EXTENSIONS OFF)
+
 # Define included source files
 set(CPP_FILES csrc/cpu_ops.cpp csrc/pythonInterface.cpp)
 set(GPU_FILES csrc/ops.cu csrc/kernels.cu)
@@ -115,8 +119,6 @@ endif()
 
 
 if (BUILD_CPU)
-    set(CMAKE_CXX_STANDARD 17)
-    set(CMAKE_CXX_STANDARD_REQUIRED ON)
     string(TOLOWER "${CMAKE_SYSTEM_PROCESSOR}" HOST_ARCH)
     if(MSVC)
         # Use the experimental OpenMP runtime for persistent thread pool support.
@@ -127,16 +129,28 @@ if (BUILD_CPU)
 endif()
 
 if(BUILD_CUDA)
+    set(CMAKE_CUDA_STANDARD 17)
+    set(CMAKE_CUDA_STANDARD_REQUIRED ON)
+    set(CMAKE_CUDA_EXTENSIONS OFF)
+
     # NVCC normally will only work with MSVC up to 1939. VS2022 17.10+ starts using versions 1940+.
     # Workaround: use --allow-unsupported-compiler
     # This needs to be added *before* we try to enable the CUDA language so CMake's compiler check passes.
+
     if(MSVC AND MSVC_VERSION VERSION_GREATER_EQUAL 1940)
         string(APPEND CMAKE_CUDA_FLAGS " --allow-unsupported-compiler")
 
-        # This is needed to build with VS2022 17.11+ and CUDA < 12.4.
+        # Suppress MSVC STL version mismatch errors when using a newer compiler than CUDA officially supports.
         if (MSVC_VERSION VERSION_GREATER_EQUAL 1941)
             string(APPEND CMAKE_CUDA_FLAGS " -D_ALLOW_COMPILER_AND_STL_VERSION_MISMATCH")
         endif()
+
+        # VS2026 (MSVC 1950+): may need _ENABLE_EXTENDED_ALIGNED_STORAGE if CUDA headers
+        # use aligned_storage with extended alignment. This is independent of C++ standard.
+        # Commented out to verify if CMAKE_CUDA_STANDARD 17 alone is sufficient.
+        # if (MSVC_VERSION VERSION_GREATER_EQUAL 1950)
+        #     string(APPEND CMAKE_CUDA_FLAGS " -D_ENABLE_EXTENDED_ALIGNED_STORAGE")
+        # endif()
     endif()
 
     enable_language(CUDA) # This will fail if CUDA is not found

From 562fc4e021d34e8b8cd489767494c6a0c73fae0d Mon Sep 17 00:00:00 2001
From: Matthew Douglas <38992547+matthewdouglas@users.noreply.github.com>
Date: Mon, 15 Jun 2026 12:59:37 -0400
Subject: [PATCH 02/11] Try improving VS2026 + CUDA < 13.2 compat.

---
 CMakeLists.txt | 9 +--------
 1 file changed, 1 insertion(+), 8 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index ccf0a5938..04175d47b 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -136,21 +136,14 @@ if(BUILD_CUDA)
     # NVCC normally will only work with MSVC up to 1939. VS2022 17.10+ starts using versions 1940+.
     # Workaround: use --allow-unsupported-compiler
     # This needs to be added *before* we try to enable the CUDA language so CMake's compiler check passes.
-
     if(MSVC AND MSVC_VERSION VERSION_GREATER_EQUAL 1940)
         string(APPEND CMAKE_CUDA_FLAGS " --allow-unsupported-compiler")
 
         # Suppress MSVC STL version mismatch errors when using a newer compiler than CUDA officially supports.
+        # This is needed to build with VS2022 17.11+ and CUDA < 12.4.
         if (MSVC_VERSION VERSION_GREATER_EQUAL 1941)
             string(APPEND CMAKE_CUDA_FLAGS " -D_ALLOW_COMPILER_AND_STL_VERSION_MISMATCH")
         endif()
-
-        # VS2026 (MSVC 1950+): may need _ENABLE_EXTENDED_ALIGNED_STORAGE if CUDA headers
-        # use aligned_storage with extended alignment. This is independent of C++ standard.
-        # Commented out to verify if CMAKE_CUDA_STANDARD 17 alone is sufficient.
-        # if (MSVC_VERSION VERSION_GREATER_EQUAL 1950)
-        #     string(APPEND CMAKE_CUDA_FLAGS " -D_ENABLE_EXTENDED_ALIGNED_STORAGE")
-        # endif()
     endif()
 
     enable_language(CUDA) # This will fail if CUDA is not found

From 3346014e6686e1a5a395c1a0c4565b4c4350defe Mon Sep 17 00:00:00 2001
From: Matthew Douglas <38992547+matthewdouglas@users.noreply.github.com>
Date: Mon, 15 Jun 2026 13:15:41 -0400
Subject: [PATCH 03/11] Try again

---
 CMakeLists.txt | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 04175d47b..2e8adf099 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -144,6 +144,10 @@ if(BUILD_CUDA)
         if (MSVC_VERSION VERSION_GREATER_EQUAL 1941)
             string(APPEND CMAKE_CUDA_FLAGS " -D_ALLOW_COMPILER_AND_STL_VERSION_MISMATCH")
         endif()
+
+        if (MSVC_VERSION VERSION_GREATER_EQUAL 1950)
+            string(APPEND CMAKE_CUDA_FLAGS " -D_ENABLE_EXTENDED_ALIGNED_STORAGE")
+        endif()
     endif()
 
     enable_language(CUDA) # This will fail if CUDA is not found

From a2fa57503aee19bf9c2cf6b9f925f4a9c29e0c60 Mon Sep 17 00:00:00 2001
From: Matthew Douglas <38992547+matthewdouglas@users.noreply.github.com>
Date: Tue, 16 Jun 2026 16:24:38 -0400
Subject: [PATCH 04/11] Downgrade to VS2022 for CUDA builds

---
 .github/scripts/build-cpu.sh         | 12 ++++-----
 .github/scripts/build-cuda.sh        | 32 +++++++++++------------
 .github/scripts/build-rocm.sh        | 30 ++++++++++-----------
 .github/scripts/build-xpu.sh         | 10 +++----
 .github/workflows/python-package.yml | 39 ++++++++++------------------
 .github/workflows/test-runner.yml    | 12 +++------
 CMakeLists.txt                       |  3 ---
 7 files changed, 59 insertions(+), 79 deletions(-)

diff --git a/.github/scripts/build-cpu.sh b/.github/scripts/build-cpu.sh
index 5db76ecce..0ede8503c 100644
--- a/.github/scripts/build-cpu.sh
+++ b/.github/scripts/build-cpu.sh
@@ -1,22 +1,22 @@
 #!/bin/bash
-declare build_arch
-declare build_os
-
 set -xeuo pipefail
 
-if [[ "${build_os}" == windows* ]]; then
+: "${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 [ "${build_os:0:5}" == macos ] && [ "${build_arch}" == aarch64 ]; then
+if [ "${RUNNER_OS}" == "macOS" ] && [ "${RUNNER_ARCH}" == "ARM64" ]; then
 	cmake -DCMAKE_OSX_ARCHITECTURES=arm64 -DCOMPUTE_BACKEND=cpu .
 else
 	cmake -DCOMPUTE_BACKEND=cpu .
 fi
 cmake --build . --config Release
 
-output_dir="output/${build_os}/${build_arch}"
+output_dir="output/${RUNNER_OS}/${RUNNER_ARCH}"
 mkdir -p "${output_dir}"
 (shopt -s nullglob && cp bitsandbytes/*.{so,dylib,dll} "${output_dir}")
diff --git a/.github/scripts/build-cuda.sh b/.github/scripts/build-cuda.sh
index 9bb2f1a66..7537ba280 100644
--- a/.github/scripts/build-cuda.sh
+++ b/.github/scripts/build-cuda.sh
@@ -1,37 +1,36 @@
 #!/bin/bash
-declare build_arch
-declare build_os
-declare cuda_version
-declare cuda_targets
-
 set -xeuo pipefail
 
-if [[ -v cuda_targets ]]; then
-    build_capability="${cuda_targets}"
-elif [ "${build_arch}" = "aarch64" ]; then
+: "${RUNNER_OS:?RUNNER_OS must be set (Linux/Windows/macOS)}"
+: "${RUNNER_ARCH:?RUNNER_ARCH must be set (X64/ARM64)}"
+: "${CUDA_VERSION:?CUDA_VERSION must be set}"
+
+if [[ -v CUDA_TARGETS ]]; then
+    build_capability="${CUDA_TARGETS}"
+elif [ "${RUNNER_ARCH}" = "ARM64" ]; then
     build_capability="75;80;90"
 
     # CUDA 12.8-12.9: Add sm100/sm120
-    [[ "${cuda_version}" == 12.8.* || "${cuda_version}" == 12.9.* ]] && build_capability="75;80;90;100;120"
+    [[ "${CUDA_VERSION}" == 12.8.* || "${CUDA_VERSION}" == 12.9.* ]] && build_capability="75;80;90;100;120"
 
     # CUDA 13.0+: Add sm100/sm110/sm120
-    [[ "${cuda_version}" == 13.*.* ]] && build_capability="75;80;90;100;110;120;121"
+    [[ "${CUDA_VERSION}" == 13.*.* ]] && build_capability="75;80;90;100;110;120;121"
 else
     # By default, target Pascal through Hopper.
     build_capability="60;70;75;80;86;89;90"
 
     # CUDA 12.8+: Add sm100 and sm120; remove < sm70 to align with PyTorch 2.8+cu128 minimum
-    [[ "${cuda_version}" == 12.8.* || "${cuda_version}" == 12.9.* ]] && build_capability="70;75;80;86;89;90;100;120"
+    [[ "${CUDA_VERSION}" == 12.8.* || "${CUDA_VERSION}" == 12.9.* ]] && build_capability="70;75;80;86;89;90;100;120"
 
     # CUDA 13.0+: Remove < sm75 to align with PyTorch 2.9+cu130 minimum
-    [[ "${cuda_version}" == 13.*.* ]] && build_capability="75;80;86;89;90;100;120"
+    [[ "${CUDA_VERSION}" == 13.*.* ]] && build_capability="75;80;86;89;90;100;120"
 fi
 
-[[ "${build_os}" = windows-* ]] && python3 -m pip install ninja
+[[ "${RUNNER_OS}" == "Windows" ]] && python3 -m pip install ninja
 
-if [ "${build_os:0:6}" == ubuntu ]; then
+if [ "${RUNNER_OS}" == "Linux" ]; then
     # We'll use Rocky Linux 8 in order to maintain manylinux 2.24 compatibility.
-    image="nvidia/cuda:${cuda_version}-devel-rockylinux8"
+    image="nvidia/cuda:${CUDA_VERSION}-devel-rockylinux8"
     echo "Using image $image"
 
     docker run -i -w /src -v "$PWD:/src" "$image" bash -c \
@@ -46,7 +45,6 @@ else
     cmake --build . --config Release
 fi
 
-
-output_dir="output/${build_os}/${build_arch}"
+output_dir="output/${RUNNER_OS}/${RUNNER_ARCH}"
 mkdir -p "${output_dir}"
 (shopt -s nullglob && cp bitsandbytes/*.{so,dylib,dll} "${output_dir}")
diff --git a/.github/scripts/build-rocm.sh b/.github/scripts/build-rocm.sh
index 77bd2eaf5..782b740c1 100644
--- a/.github/scripts/build-rocm.sh
+++ b/.github/scripts/build-rocm.sh
@@ -1,21 +1,21 @@
 #!/bin/bash
-declare build_arch
-declare build_os
-declare rocm_version
-
 set -xeuo pipefail
+
+: "${RUNNER_OS:?RUNNER_OS must be set (Linux/Windows)}"
+: "${ROCM_VERSION:?ROCM_VERSION must be set}"
+
 bnb_rocm_arch="gfx90a;gfx942;gfx1100;gfx1101;gfx1102;gfx1103"
 
 # ROCm 6.4+ - Add RDNA4 and RDNA3.5 targets. Note we assume >=6.4.4.
-[[ "${rocm_version}" == 6.4.* || "${rocm_version}" == 7.* ]] && bnb_rocm_arch="${bnb_rocm_arch};gfx1150;gfx1151;gfx1152;gfx1153;gfx1200;gfx1201"
+[[ "${ROCM_VERSION}" == 6.4.* || "${ROCM_VERSION}" == 7.* ]] && bnb_rocm_arch="${bnb_rocm_arch};gfx1150;gfx1151;gfx1152;gfx1153;gfx1200;gfx1201"
 
 # ROCm 7.0+ - Add gfx950
-[[ "${rocm_version}" == 7.* ]] && bnb_rocm_arch="${bnb_rocm_arch};gfx950"
+[[ "${ROCM_VERSION}" == 7.* ]] && bnb_rocm_arch="${bnb_rocm_arch};gfx950"
 
-if [ "${build_os:0:6}" == ubuntu ]; then
-    image=rocm/dev-ubuntu-22.04:${rocm_version}-complete
+if [ "${RUNNER_OS}" == "Linux" ]; then
+    image=rocm/dev-ubuntu-22.04:${ROCM_VERSION}-complete
     echo "Using image $image"
-    docker run --rm --platform "linux/$build_arch" -i \
+    docker run --rm -i \
         -w /src -v "$PWD:/src" "$image" sh -c \
         "apt-get update \
       && pip install cmake==3.31.6 \
@@ -27,12 +27,12 @@ else
     pip install ninja cmake==3.31.6
 
     # Install ROCm SDK wheels from repo.radeon.com.
-    rocm_base_url="https://repo.radeon.com/rocm/windows/rocm-rel-${rocm_version}"
+    rocm_base_url="https://repo.radeon.com/rocm/windows/rocm-rel-${ROCM_VERSION}"
     pip install \
-        "${rocm_base_url}/rocm_sdk_core-${rocm_version}-py3-none-win_amd64.whl" \
-        "${rocm_base_url}/rocm_sdk_devel-${rocm_version}-py3-none-win_amd64.whl" \
-        "${rocm_base_url}/rocm_sdk_libraries_custom-${rocm_version}-py3-none-win_amd64.whl" \
-        "${rocm_base_url}/rocm-${rocm_version}.tar.gz"
+        "${rocm_base_url}/rocm_sdk_core-${ROCM_VERSION}-py3-none-win_amd64.whl" \
+        "${rocm_base_url}/rocm_sdk_devel-${ROCM_VERSION}-py3-none-win_amd64.whl" \
+        "${rocm_base_url}/rocm_sdk_libraries_custom-${ROCM_VERSION}-py3-none-win_amd64.whl" \
+        "${rocm_base_url}/rocm-${ROCM_VERSION}.tar.gz"
 
     # Expand the devel tarball
     rocm-sdk init
@@ -50,6 +50,6 @@ else
     cmake --build .
 fi
 
-output_dir="output/${build_os}/${build_arch}"
+output_dir="output/${RUNNER_OS}/X64"
 mkdir -p "${output_dir}"
 (shopt -s nullglob && cp bitsandbytes/*.{so,dylib,dll} "${output_dir}")
diff --git a/.github/scripts/build-xpu.sh b/.github/scripts/build-xpu.sh
index d069e1230..9c9e51e41 100755
--- a/.github/scripts/build-xpu.sh
+++ b/.github/scripts/build-xpu.sh
@@ -1,10 +1,10 @@
 #!/bin/bash
-declare build_os
-
 set -xeuo pipefail
 
-# We currently only build XPU on Linux.
-if [ "${build_os:0:6}" == ubuntu ]; then
+: "${RUNNER_OS:?RUNNER_OS must be set (Linux/Windows)}"
+
+# We currently only build XPU on Linux x64 and Windows x64.
+if [ "${RUNNER_OS}" == "Linux" ]; then
     # TODO: We might want to pre-build this as our own customized image in the future.
     image=intel/deep-learning-essentials:2025.1.3-0-devel-ubuntu22.04
     echo "Using image $image"
@@ -17,6 +17,6 @@ if [ "${build_os:0:6}" == ubuntu ]; then
       && cmake --build . --config Release"
 fi
 
-output_dir="output/${build_os}/x86_64"
+output_dir="output/${RUNNER_OS}/X64"
 mkdir -p "${output_dir}"
 (shopt -s nullglob && cp bitsandbytes/*.{so,dylib,dll} "${output_dir}")
diff --git a/.github/workflows/python-package.yml b/.github/workflows/python-package.yml
index 0e88ba018..4806007a9 100644
--- a/.github/workflows/python-package.yml
+++ b/.github/workflows/python-package.yml
@@ -52,13 +52,10 @@ jobs:
           arch: ${{ matrix.arch == 'arm64' && 'arm64' || 'x64' }}
       - name: Build C++
         run: bash .github/scripts/build-cpu.sh
-        env:
-          build_os: ${{ matrix.os }}
-          build_arch: ${{ matrix.arch }}
       - name: Upload build artifact
         uses: actions/upload-artifact@v4
         with:
-          name: shared_library_${{ matrix.os }}_${{ matrix.arch }}
+          name: shared_library_${{ runner.os }}_${{ runner.arch }}
           path: output/*
           retention-days: 7
 
@@ -69,13 +66,13 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        os: [ubuntu-22.04, ubuntu-22.04-arm, windows-2025]
+        os: [ubuntu-22.04, ubuntu-22.04-arm, windows-2022]
         include:
           - os: ubuntu-22.04
             arch: x86_64
           - os: ubuntu-22.04-arm
             arch: aarch64
-          - os: windows-2025
+          - os: windows-2022
             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", "12.9.1", "13.0.2", "13.2.0"]
@@ -93,45 +90,39 @@ jobs:
           sub-packages: ${{ format('["nvcc"{0},"cudart","cublas","thrust","cublas_dev"]', startsWith(matrix.cuda_version, '13.') && ',"crt","nvvm","nvptxcompiler"' || '') }}
           use-github-cache: false
           use-local-cache: false
-          log-file-suffix: ${{matrix.os}}-${{matrix.cuda_version}}.txt
+          log-file-suffix: ${{ runner.os }}-${{ runner.arch }}-${{matrix.cuda_version}}.txt
       - name: Setup MSVC
         if: startsWith(matrix.os, 'windows')
         uses: ilammy/msvc-dev-cmd@v1.13.0 # to use cl
       - name: Build C++
         run: bash .github/scripts/build-cuda.sh
         env:
-          build_os: ${{ matrix.os }}
-          build_arch: ${{ matrix.arch }}
-          cuda_version: ${{ matrix.cuda_version }}
+          CUDA_VERSION: ${{ matrix.cuda_version }}
       - name: Upload build artifact
         uses: actions/upload-artifact@v4
         with:
-          name: shared_library_cuda_${{ matrix.os }}_${{ matrix.arch }}_${{ matrix.cuda_version }}
+          name: shared_library_cuda_${{ runner.os }}_${{ runner.arch }}_${{ matrix.cuda_version }}
           path: output/*
           retention-days: 7
 
   build-xpu:
     strategy:
       matrix:
-        os: [ubuntu-22.04, windows-2025]
+        os: [ubuntu-22.04, windows-2022]
     runs-on: ${{ matrix.os }}
     steps:
       - uses: actions/checkout@v4
       - name: Build C++ (Linux)
         if: runner.os == 'Linux'
         run: bash .github/scripts/build-xpu.sh
-        env:
-          build_os: ${{ matrix.os }}
       - name: Build C++ (Windows)
         if: runner.os == 'Windows'
         run: .github/scripts/build-xpu-windows.bat
         shell: cmd
-        env:
-          build_os: ${{ matrix.os }}
       - name: Upload build artifact
         uses: actions/upload-artifact@v4
         with:
-          name: shared_library_xpu_${{ matrix.os }}_x86_64
+          name: shared_library_xpu_${{ runner.os }}_${{ runner.arch }}
           path: output/*
           retention-days: 7
 
@@ -171,13 +162,11 @@ jobs:
       - name: Build C++
         run: bash .github/scripts/build-rocm.sh
         env:
-          build_os: ${{ matrix.os }}
-          build_arch: ${{ matrix.arch }}
-          rocm_version: ${{ matrix.rocm_version }}
+          ROCM_VERSION: ${{ matrix.rocm_version }}
       - name: Upload build artifact
         uses: actions/upload-artifact@v4
         with:
-          name: shared_library_rocm_${{ matrix.os }}_${{ matrix.arch }}_${{ matrix.rocm_version }}
+          name: shared_library_rocm_${{ runner.os }}_${{ runner.arch }}_${{ matrix.rocm_version }}
           path: output/*
           retention-days: 7
 
@@ -220,13 +209,13 @@ jobs:
         uses: actions/download-artifact@v4
         with:
           merge-multiple: true
-          pattern: "shared_library*_${{ matrix.os }}_${{ matrix.arch }}*"
+          pattern: "shared_library*_${{ runner.os }}_${{ runner.arch }}*"
           path: output/
       - name: Copy correct platform shared library
         shell: bash
         run: |
           ls -lR output/
-          cp output/${{ matrix.os }}/${{ matrix.arch }}/* bitsandbytes/
+          cp output/${{ runner.os }}/${{ runner.arch }}/* bitsandbytes/
       - name: Set up Python ${{ matrix.python-version }}
         uses: actions/setup-python@v5
         with:
@@ -243,7 +232,7 @@ jobs:
       - name: Upload build artifact
         uses: actions/upload-artifact@v4
         with:
-          name: bdist_wheel_${{ matrix.os }}_${{ matrix.arch }}
+          name: bdist_wheel_${{ runner.os }}_${{ runner.arch }}
           path: dist/bitsandbytes-*.whl
           retention-days: 7
 
@@ -406,7 +395,7 @@ jobs:
       - name: Download wheel
         uses: actions/download-artifact@v4
         with:
-          name: bdist_wheel_${{ matrix.os }}_${{ matrix.arch }}
+          name: bdist_wheel_${{ runner.os }}_${{ runner.arch }}
           path: wheels/
       - name: Set up Python
         uses: actions/setup-python@v5
diff --git a/.github/workflows/test-runner.yml b/.github/workflows/test-runner.yml
index d1e81bc32..db1414375 100644
--- a/.github/workflows/test-runner.yml
+++ b/.github/workflows/test-runner.yml
@@ -43,6 +43,7 @@ jobs:
         inputs.platform == 'linux-aarch64' && 'ubuntu-22.04-arm' ||
         inputs.platform == 'macos' && 'macos-15' ||
         inputs.platform == 'windows-arm64' && 'windows-11-arm' ||
+        (inputs.platform == 'windows' && inputs.backend == 'cuda') && 'windows-2022' ||
         'windows-2025'
       }}
     outputs:
@@ -140,11 +141,6 @@ jobs:
 
       - uses: actions/checkout@v4
 
-      - name: Set build environment variables
-        shell: bash
-        run: |
-          echo "build_os=${{ steps.config.outputs.build_os }}" >> $GITHUB_ENV
-          echo "build_arch=${{ steps.config.outputs.arch }}" >> $GITHUB_ENV
 
       # Windows + CUDA: Install CUDA Toolkit
       - name: Install CUDA Toolkit
@@ -173,14 +169,14 @@ jobs:
         if: inputs.backend == 'cuda'
         run: bash .github/scripts/build-cuda.sh
         env:
-          cuda_version: ${{ inputs.cuda_version }}
-          cuda_targets: "75;80;89"
+          CUDA_VERSION: ${{ inputs.cuda_version }}
+          CUDA_TARGETS: "75;80;89"
 
       - name: Upload build artifact
         uses: actions/upload-artifact@v4
         with:
           name: ${{ steps.config.outputs.artifact_name }}
-          path: output/${{ steps.config.outputs.build_os }}/${{ steps.config.outputs.arch }}/*
+          path: output/${{ runner.os }}/${{ runner.arch }}/*
           retention-days: 7
 
   test:
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 2e8adf099..fc7e41aac 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -145,9 +145,6 @@ if(BUILD_CUDA)
             string(APPEND CMAKE_CUDA_FLAGS " -D_ALLOW_COMPILER_AND_STL_VERSION_MISMATCH")
         endif()
 
-        if (MSVC_VERSION VERSION_GREATER_EQUAL 1950)
-            string(APPEND CMAKE_CUDA_FLAGS " -D_ENABLE_EXTENDED_ALIGNED_STORAGE")
-        endif()
     endif()
 
     enable_language(CUDA) # This will fail if CUDA is not found

From 363df921b21b1e7404e3eb5981822397fd297866 Mon Sep 17 00:00:00 2001
From: Matthew Douglas <38992547+matthewdouglas@users.noreply.github.com>
Date: Tue, 16 Jun 2026 17:22:59 -0400
Subject: [PATCH 05/11] Update windows ROCm/XPU builds

---
 .github/scripts/build-rocm.sh         | 2 --
 .github/scripts/build-xpu-windows.bat | 2 +-
 2 files changed, 1 insertion(+), 3 deletions(-)

diff --git a/.github/scripts/build-rocm.sh b/.github/scripts/build-rocm.sh
index 782b740c1..d1bdd9765 100644
--- a/.github/scripts/build-rocm.sh
+++ b/.github/scripts/build-rocm.sh
@@ -24,8 +24,6 @@ if [ "${RUNNER_OS}" == "Linux" ]; then
 else
     bnb_rocm_arch="gfx1100;gfx1101;gfx1102;gfx1150;gfx1151;gfx1200;gfx1201"
 
-    pip install ninja cmake==3.31.6
-
     # Install ROCm SDK wheels from repo.radeon.com.
     rocm_base_url="https://repo.radeon.com/rocm/windows/rocm-rel-${ROCM_VERSION}"
     pip install \
diff --git a/.github/scripts/build-xpu-windows.bat b/.github/scripts/build-xpu-windows.bat
index c7317b8a7..02d281d6e 100644
--- a/.github/scripts/build-xpu-windows.bat
+++ b/.github/scripts/build-xpu-windows.bat
@@ -29,6 +29,6 @@ if ERRORLEVEL 1 (
 )
 echo ::endgroup::
 
-set output_dir=output\%build_os%\x86_64
+set output_dir=output\Windows\X64
 if not exist "%output_dir%" mkdir "%output_dir%"
 copy bitsandbytes\*.dll "%output_dir%\" 2>nul

From c4beacbf4bd659ec6f02d1ab8544f7495f187d46 Mon Sep 17 00:00:00 2001
From: Matthew Douglas <38992547+matthewdouglas@users.noreply.github.com>
Date: Tue, 16 Jun 2026 18:11:08 -0400
Subject: [PATCH 06/11] update rocm build

---
 .github/scripts/build-rocm.sh | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/.github/scripts/build-rocm.sh b/.github/scripts/build-rocm.sh
index d1bdd9765..9d452b8bc 100644
--- a/.github/scripts/build-rocm.sh
+++ b/.github/scripts/build-rocm.sh
@@ -35,15 +35,15 @@ else
     # Expand the devel tarball
     rocm-sdk init
 
-    ROCM_PATH="$(rocm-sdk path --root)"
-    export ROCM_PATH
-    export PATH="${ROCM_PATH}/bin:${PATH}"
+    ROCM_PATH="$(rocm-sdk path --root | tr '\\' '/')"
+    export ROCM_PATH PATH="${ROCM_PATH}/bin:${PATH}"
 
     cmake -G Ninja \
         -DCOMPUTE_BACKEND=hip \
         -DBNB_ROCM_ARCH="${bnb_rocm_arch}" \
         -DCMAKE_BUILD_TYPE=MinSizeRel \
         -DCMAKE_HIP_FLAGS="--offload-compress" \
+        -DCMAKE_HIP_COMPILER_ROCM_ROOT="${ROCM_PATH}" \
         -S .
     cmake --build .
 fi

From e40e8edef0c57a62e1a6797c538725eb56e61182 Mon Sep 17 00:00:00 2001
From: Matthew Douglas <38992547+matthewdouglas@users.noreply.github.com>
Date: Tue, 16 Jun 2026 18:26:19 -0400
Subject: [PATCH 07/11] update windows rocm

---
 .github/workflows/python-package.yml | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/.github/workflows/python-package.yml b/.github/workflows/python-package.yml
index 4806007a9..dcebbe907 100644
--- a/.github/workflows/python-package.yml
+++ b/.github/workflows/python-package.yml
@@ -159,6 +159,8 @@ jobs:
       - name: Setup MSVC
         if: startsWith(matrix.os, 'windows')
         uses: ilammy/msvc-dev-cmd@v1.13.0
+        with:
+          toolset: "14.44"
       - name: Build C++
         run: bash .github/scripts/build-rocm.sh
         env:

From f8c4e5b9a025b6f3e828106cfa6ff98ed1244d69 Mon Sep 17 00:00:00 2001
From: Matthew Douglas <38992547+matthewdouglas@users.noreply.github.com>
Date: Wed, 17 Jun 2026 13:57:53 -0400
Subject: [PATCH 08/11] try cuda windows-2025 again

---
 .github/workflows/python-package.yml | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/.github/workflows/python-package.yml b/.github/workflows/python-package.yml
index dcebbe907..2dfb9654d 100644
--- a/.github/workflows/python-package.yml
+++ b/.github/workflows/python-package.yml
@@ -66,13 +66,13 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        os: [ubuntu-22.04, ubuntu-22.04-arm, windows-2022]
+        os: [ubuntu-22.04, ubuntu-22.04-arm, windows-2025]
         include:
           - os: ubuntu-22.04
             arch: x86_64
           - os: ubuntu-22.04-arm
             arch: aarch64
-          - os: windows-2022
+          - os: windows-2025
             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", "12.9.1", "13.0.2", "13.2.0"]
@@ -94,6 +94,8 @@ jobs:
       - name: Setup MSVC
         if: startsWith(matrix.os, 'windows')
         uses: ilammy/msvc-dev-cmd@v1.13.0 # to use cl
+        with:
+          toolset: "14.44"
       - name: Build C++
         run: bash .github/scripts/build-cuda.sh
         env:

From 1756d7d6f1130d94a2e59fe25d558842ad39b116 Mon Sep 17 00:00:00 2001
From: Matthew Douglas <38992547+matthewdouglas@users.noreply.github.com>
Date: Wed, 17 Jun 2026 14:57:16 -0400
Subject: [PATCH 09/11] update doc

---
 docs/source/installation.mdx | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)

diff --git a/docs/source/installation.mdx b/docs/source/installation.mdx
index 9a335f2d8..1a706dee4 100644
--- a/docs/source/installation.mdx
+++ b/docs/source/installation.mdx
@@ -174,7 +174,7 @@ The currently distributed `bitsandbytes` packages are built with the following c
 |--------------------|----------------------|----------------------|
 | **Linux x86-64**   | GCC 11.4             | AVX2                 |
 | **Linux aarch64**  | GCC 11.4             |                      |
-| **Windows x86-64** | MSVC 19.43+ (VS2022) | AVX2                 |
+| **Windows x86-64** | MSVC 19.51+ (VS2026) | AVX2                 |
 | **Windows arm64**  | MSVC 19.43+ (VS2022) | ARM NEON             |
 | **macOS arm64**    | Apple Clang 17       |                      |
 
@@ -201,6 +201,8 @@ pip install -e .
 </hfoption>
 <hfoption id="Windows (x86-64)">
 
+Requires Visual Studio 2022 or 2026.
+
 ```bash
 git clone https://github.com/bitsandbytes-foundation/bitsandbytes.git && cd bitsandbytes/
 pip install -e .
@@ -209,7 +211,7 @@ pip install -e .
 </hfoption>
 <hfoption id="Windows (ARM64)">
 
-Requires Visual Studio 2022 with the **ARM64 C++ build tools** component, Python >= **3.12**, and PyTorch >= **2.12**.
+Requires Visual Studio 2022 or 2026 with the **ARM64 C++ build tools** component, Python >= **3.12**, and PyTorch >= **2.12**.
 
 ```bash
 git clone https://github.com/bitsandbytes-foundation/bitsandbytes.git && cd bitsandbytes/
@@ -280,7 +282,7 @@ pip install -e .   # `-e` for "editable" install, when developing BNB (otherwise
 </hfoption>
 <hfoption id="Windows">
 
-Compilation on Windows requires Visual Studio with C++ support, CMake, Ninja, and Python >= **3.10**.
+Compilation on Windows requires Visual Studio 2022 with C++ support, CMake, Ninja, and Python >= **3.10**.
 
 Instead of a system-wide ROCm installation, you can use the pip-installable ROCm SDK wheels from [repo.radeon.com](https://repo.radeon.com/rocm/windows/):
 

From bc9993fd539016003e6e6ce10631b001010ebfe0 Mon Sep 17 00:00:00 2001
From: Matthew Douglas <38992547+matthewdouglas@users.noreply.github.com>
Date: Wed, 17 Jun 2026 14:57:32 -0400
Subject: [PATCH 10/11] update test workflow

---
 .github/workflows/test-runner.yml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.github/workflows/test-runner.yml b/.github/workflows/test-runner.yml
index db1414375..d64e66639 100644
--- a/.github/workflows/test-runner.yml
+++ b/.github/workflows/test-runner.yml
@@ -43,7 +43,6 @@ jobs:
         inputs.platform == 'linux-aarch64' && 'ubuntu-22.04-arm' ||
         inputs.platform == 'macos' && 'macos-15' ||
         inputs.platform == 'windows-arm64' && 'windows-11-arm' ||
-        (inputs.platform == 'windows' && inputs.backend == 'cuda') && 'windows-2022' ||
         'windows-2025'
       }}
     outputs:
@@ -158,6 +157,7 @@ jobs:
         uses: ilammy/msvc-dev-cmd@v1.13.0
         with:
           arch: ${{ inputs.platform == 'windows-arm64' && 'arm64' || 'x64' }}
+          toolset: ${{ (inputs.platform == 'windows' && inputs.backend == 'cuda') && '14.44' || '' }}
 
       # Build CPU backend
       - name: Build C++

From 96771e2a19ddbf62ebdcbee3692410dad89eb626 Mon Sep 17 00:00:00 2001
From: Matthew Douglas <38992547+matthewdouglas@users.noreply.github.com>
Date: Wed, 17 Jun 2026 17:07:24 -0400
Subject: [PATCH 11/11] cleanup

---
 .github/scripts/set_platform_tag.py  |  8 +++--
 .github/workflows/python-package.yml | 52 ++++------------------------
 .github/workflows/test-runner.yml    |  4 ---
 3 files changed, 11 insertions(+), 53 deletions(-)

diff --git a/.github/scripts/set_platform_tag.py b/.github/scripts/set_platform_tag.py
index 1ffeeec9d..0186d41c4 100644
--- a/.github/scripts/set_platform_tag.py
+++ b/.github/scripts/set_platform_tag.py
@@ -4,14 +4,16 @@
 
 
 def get_platform_tag(architecture):
+    arch = architecture.lower()
+    is_x64 = arch in ("x86_64", "x64")
     system = platform.system()
 
     if system == "Linux":
-        tag = "manylinux_2_24_x86_64" if architecture == "x86_64" else "manylinux_2_24_aarch64"
+        tag = "manylinux_2_24_x86_64" if is_x64 else "manylinux_2_24_aarch64"
     elif system == "Darwin":
         tag = "macosx_14_0_arm64"
     elif system == "Windows":
-        tag = "win_amd64" if architecture == "x86_64" else "win_arm64"
+        tag = "win_amd64" if is_x64 else "win_arm64"
     else:
         sys.exit(f"Unsupported system: {system}")
 
@@ -20,7 +22,7 @@ def get_platform_tag(architecture):
 
 def main():
     parser = argparse.ArgumentParser(description="Determine platform tag.")
-    parser.add_argument("arch", type=str, help="Architecture (e.g., x86_64, aarch64)")
+    parser.add_argument("arch", type=str, help="Architecture (e.g., x86_64, aarch64, X64, ARM64)")
     args = parser.parse_args()
 
     tag = get_platform_tag(args.arch)
diff --git a/.github/workflows/python-package.yml b/.github/workflows/python-package.yml
index 2dfb9654d..2d0e7ca4b 100644
--- a/.github/workflows/python-package.yml
+++ b/.github/workflows/python-package.yml
@@ -31,17 +31,7 @@ jobs:
   build-cpu:
     strategy:
       matrix:
-        include:
-          - os: ubuntu-22.04
-            arch: x86_64
-          - os: ubuntu-22.04-arm
-            arch: aarch64
-          - os: windows-2025
-            arch: x86_64
-          - os: windows-11-arm
-            arch: arm64
-          - os: macos-15
-            arch: arm64
+        os: [ubuntu-22.04, ubuntu-22.04-arm, windows-2025, windows-11-arm, macos-15]
     runs-on: ${{ matrix.os }}
     steps:
       - uses: actions/checkout@v4
@@ -49,7 +39,7 @@ jobs:
         if: startsWith(matrix.os, 'windows')
         uses: ilammy/msvc-dev-cmd@v1.13.0 # to use cl
         with:
-          arch: ${{ matrix.arch == 'arm64' && 'arm64' || 'x64' }}
+          arch: ${{ runner.arch == 'ARM64' && 'arm64' || 'x64' }}
       - name: Build C++
         run: bash .github/scripts/build-cpu.sh
       - name: Upload build artifact
@@ -67,13 +57,6 @@ jobs:
       fail-fast: false
       matrix:
         os: [ubuntu-22.04, ubuntu-22.04-arm, windows-2025]
-        include:
-          - os: ubuntu-22.04
-            arch: x86_64
-          - os: ubuntu-22.04-arm
-            arch: aarch64
-          - os: windows-2025
-            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", "12.9.1", "13.0.2", "13.2.0"]
     runs-on: ${{ matrix.os }}
@@ -132,11 +115,9 @@ jobs:
     strategy:
       matrix:
         os: [ubuntu-22.04]
-        arch: [x86_64]
         rocm_version: ["6.2.4", "6.3.4", "6.4.4", "7.0.2", "7.1.1", "7.2.3"]
         include:
           - os: windows-2025
-            arch: x86_64
             rocm_version: "7.2.1"
     runs-on: ${{ matrix.os }}
     steps:
@@ -186,23 +167,6 @@ jobs:
     strategy:
       matrix:
         os: [ubuntu-22.04, ubuntu-22.04-arm, windows-2025, windows-11-arm, macos-15]
-        include:
-          - os: ubuntu-22.04
-            arch: x86_64
-            python-version: "3.10"
-          - os: ubuntu-22.04-arm
-            arch: aarch64
-            python-version: "3.10"
-          - os: windows-2025
-            arch: x86_64
-            python-version: "3.10"
-          - os: windows-11-arm
-            arch: arm64
-            # Python for Windows ARM64 is only available from 3.12+
-            python-version: "3.12"
-          - os: macos-15
-            arch: arm64
-            python-version: "3.10"
         # The specific Python version is irrelevant in this context as we are only packaging non-C extension
         # code. This ensures compatibility across Python versions, as compatibility is
         # dictated by the packaged code itself, not the Python version used for packaging.
@@ -220,17 +184,18 @@ jobs:
         run: |
           ls -lR output/
           cp output/${{ runner.os }}/${{ runner.arch }}/* bitsandbytes/
-      - name: Set up Python ${{ matrix.python-version }}
+      - name: Set up Python
         uses: actions/setup-python@v5
         with:
-          python-version: ${{ matrix.python-version }}
+          # Python for Windows ARM64 is only available from 3.12+
+          python-version: ${{ matrix.os == 'windows-11-arm' && '3.12' || '3.10' }}
           cache: pip
       - run: pip install build wheel
       - run: python -m build .
       - name: Determine and Set Platform Tag, then Tag Wheel
         shell: bash
         run: |
-          PLATFORM_TAG=$(python .github/scripts/set_platform_tag.py "${{ matrix.arch }}")
+          PLATFORM_TAG=$(python .github/scripts/set_platform_tag.py "${{ runner.arch }}")
           echo "PLATFORM_TAG=$PLATFORM_TAG"
           wheel tags --remove --abi-tag=none --python-tag=py3 --platform-tag=$PLATFORM_TAG dist/bitsandbytes-*.whl
       - name: Upload build artifact
@@ -386,11 +351,6 @@ jobs:
     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
diff --git a/.github/workflows/test-runner.yml b/.github/workflows/test-runner.yml
index d64e66639..717c73265 100644
--- a/.github/workflows/test-runner.yml
+++ b/.github/workflows/test-runner.yml
@@ -48,8 +48,6 @@ jobs:
     outputs:
       test_runner: ${{ steps.config.outputs.test_runner }}
       artifact_name: ${{ steps.config.outputs.artifact_name }}
-      build_os: ${{ steps.config.outputs.build_os }}
-      arch: ${{ steps.config.outputs.arch }}
     steps:
       - name: Configure test runner and paths
         id: config
@@ -135,8 +133,6 @@ jobs:
 
           echo "test_runner=${TEST_RUNNER}" >> $GITHUB_OUTPUT
           echo "artifact_name=${ARTIFACT}" >> $GITHUB_OUTPUT
-          echo "build_os=${BUILD_OS}" >> $GITHUB_OUTPUT
-          echo "arch=${ARCH}" >> $GITHUB_OUTPUT
 
       - uses: actions/checkout@v4