Merge branch 'main' into matmul4bit

Author: ved1beta

.clang-format

@@ -0,0 +1,36 @@
+---
+BasedOnStyle: LLVM
+AlignAfterOpenBracket: BlockIndent
+BinPackArguments: true
+BinPackParameters: true
+BracedInitializerIndentWidth: 4
+ColumnLimit: 120
+Cpp11BracedListStyle: true
+IndentWidth: 4
+IndentWrappedFunctionNames: true
+PointerAlignment: Left
+SeparateDefinitionBlocks: Always
+Standard: c++17
+StatementMacros:
+  - 'MAKE_PreconditionOptimizer32bit1State'
+  - 'MAKE_PreconditionOptimizer32bit2State'
+  - 'MAKE_PreconditionStatic8bit1State'
+  - 'MAKE_PreconditionStatic8bit2State'
+  - 'MAKE_Optimizer32bit1State'
+  - 'MAKE_optimizerStatic8bit1State'
+  - 'MAKE_optimizerStatic8bit2State'
+  - 'MAKE_OptimizerStatic8bit1StateBlockwise'
+  - 'MAKE_OptimizerStatic8bit2StateBlockwise'
+  - 'MAKE_kQuantizeBlockwise'
+  - 'MAKE_BLOCKWISE8'
+  - 'MAKE_ELEMENTWISE_FUNC'
+  - 'CMAKE_ELEMENTWISE_FUNC'
+  - 'MAKE_FUNC8'
+  - 'MAKE_FUNC32'
+  - 'MAKE_CBLOCKWISE8'
+  - 'MAKE_CFUNC8'
+  - 'MAKE_CFUNC32'
+
+UseTab: Never
+
+...

.git-blame-ignore-revs

@@ -15,3 +15,6 @@ ea7c14f8ef64924f2d0ff80df3cdabf2c7299848
 
 # CHANGELOG: to reverse chron order + mdformat
 4743ff0d43e04e4cc3e5d8b9e7cd016c0defa36d
+
+# Apply clang-format
+4955d136ae083c2be1236d8915913166e1790aad

.gitattributes

@@ -0,0 +1 @@
+*.bat text eol=crlf

.github/FUNDING.yml

@@ -0,0 +1 @@
+open_collective: bitsandbytes

.github/scripts/build-cuda.sh

@@ -11,14 +11,20 @@ if [[ -v cuda_targets ]]; then
 elif [ "${build_arch}" = "aarch64" ]; then
     build_capability="75;80;90"
 
-    # CUDA 12.8: Add sm100
-    [[ "${cuda_version}" == 12.8.* ]] && build_capability="75;80;90;100"
+    # CUDA 12.8-12.9: Add sm100/sm120
+    [[ "${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"
 else
-    # By default, target Maxwell through Hopper.
-    build_capability="50;52;60;61;70;75;80;86;89;90"
+    # 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 12.8: Add sm100 and sm120; remove < sm75 to align with PyTorch 2.7+cu128 minimum
-    [[ "${cuda_version}" == 12.8.* ]] && build_capability="75;80;86;89;90;100;120"
+    # CUDA 13.0+: Remove < sm75 to align with PyTorch 2.9+cu130 minimum
+    [[ "${cuda_version}" == 13.*.* ]] && build_capability="75;80;86;89;90;100;120"
 fi
 
 [[ "${build_os}" = windows-* ]] && python3 -m pip install ninja
@@ -29,8 +35,8 @@ if [ "${build_os:0:6}" == ubuntu ]; then
     echo "Using image $image"
 
     docker run -i -w /src -v "$PWD:/src" "$image" bash -c \
-        "dnf update -y \
-        && dnf install cmake gcc-toolset-11 -y \
+        "dnf -y --refresh update --security \
+        && dnf -y install cmake gcc-toolset-11 --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"

.github/scripts/build-rocm.sh

@@ -0,0 +1,28 @@
+#!/bin/bash
+declare build_arch
+declare build_os
+declare rocm_version
+
+set -xeuo pipefail
+bnb_rocm_arch="gfx90a;gfx942;gfx1100;gfx1101"
+
+# ROCm 6.4+ - Add gfx1150/gfx1151/gfx1200/gfx1201. Note we assume >=6.4.4.
+[[ "${rocm_version}" == 6.4.* || "${rocm_version}" == 7.* ]] && bnb_rocm_arch="${bnb_rocm_arch};gfx1150;gfx1151;gfx1200;gfx1201"
+
+# ROCm 7.0+ - Add gfx950
+[[ "${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
+    echo "Using image $image"
+    docker run --rm --platform "linux/$build_arch" -i \
+        -w /src -v "$PWD:/src" "$image" sh -c \
+        "apt-get update \
+      && pip install cmake==3.31.6 \
+      && cmake -DCOMPUTE_BACKEND=hip -DCMAKE_BUILD_TYPE=MinSizeRel -DCMAKE_HIP_FLAGS=\"--offload-compress\" -DBNB_ROCM_ARCH=\"${bnb_rocm_arch}\" . \
+      && cmake --build ."
+fi
+
+output_dir="output/${build_os}/${build_arch}"
+mkdir -p "${output_dir}"
+(shopt -s nullglob && cp bitsandbytes/*.{so,dylib,dll} "${output_dir}")

.github/scripts/build-xpu-windows.bat

@@ -0,0 +1,34 @@
+set INTEL_DLE_URL=https://registrationcenter-download.intel.com/akdlm/IRC_NAS/75d4eb97-914a-4a95-852c-7b9733d80f74/intel-deep-learning-essentials-2025.1.3.8_offline.exe
+set INTEL_DLE_TMP=%RUNNER_TEMP%\intel_dle
+set INTEL_DLE_LOG=%RUNNER_TEMP%\intel_dle_log.txt
+
+echo ::group::Intel Deep Learning Essentials Installation
+curl -o intel-dle-installer.exe %INTEL_DLE_URL%
+start /wait "Intel DLE Install" intel-dle-installer.exe -f %INTEL_DLE_TMP% -l %INTEL_DLE_LOG% --silent -a --eula=accept -p=NEED_VS2022_INTEGRATION=0
+type %INTEL_DLE_LOG%
+if ERRORLEVEL 1 (
+    echo Failed to install Intel Deep Learning Essentials
+    exit /b 1
+)
+echo ::endgroup::
+
+echo ::group::Build Environment Setup
+call "%ProgramFiles(x86)%\Intel\oneAPI\setvars.bat"
+cmake -G Ninja -DCOMPUTE_BACKEND=xpu -DCMAKE_BUILD_TYPE=Release .
+if ERRORLEVEL 1 (
+    echo Failed to setup environment
+    exit /b 1
+)
+echo ::endgroup::
+
+echo ::group::Building with XPU backend
+cmake --build . --config Release
+if ERRORLEVEL 1 (
+    echo Build failed
+    exit /b 1
+)
+echo ::endgroup::
+
+set output_dir=output\%build_os%\x86_64
+if not exist "%output_dir%" mkdir "%output_dir%"
+copy bitsandbytes\*.dll "%output_dir%\" 2>nul

.github/scripts/build-xpu.sh

@@ -0,0 +1,22 @@
+#!/bin/bash
+declare build_os
+
+set -xeuo pipefail
+
+# We currently only build XPU on Linux.
+if [ "${build_os:0:6}" == ubuntu ]; 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 $image"
+    docker run --rm -i \
+        -w /src -v "$PWD:/src" "$image" sh -c \
+        "apt-get update \
+      && DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
+        cmake bison intel-fw-gpu intel-ocloc \
+      && cmake -DCOMPUTE_BACKEND=xpu . \
+      && cmake --build . --config Release"
+fi
+
+output_dir="output/${build_os}/x86_64"
+mkdir -p "${output_dir}"
+(shopt -s nullglob && cp bitsandbytes/*.{so,dylib,dll} "${output_dir}")

.github/scripts/set_platform_tag.py

@@ -9,7 +9,7 @@ def get_platform_tag(architecture):
     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"
+        tag = "macosx_14_0_arm64"
     elif system == "Windows":
         tag = "win_amd64" if architecture == "x86_64" else "win_arm64"
     else: