Windows CI (#3021)

.github/actions/build-windows/action.yml

@@ -0,0 +1,18 @@
+name: 'Build on Windows'
+
+runs:
+  using: 'composite'
+  steps:
+    - name: Build CPP only
+      shell: cmd
+      env:
+        # For MSVC, Ninja/Release is the only config supported by ccache.
+        CMAKE_ARGS: >-
+          -G Ninja
+          -DCMAKE_BUILD_TYPE=Release
+          -DCMAKE_C_COMPILER=cl
+          -DCMAKE_CXX_COMPILER=cl
+          -DCMAKE_RC_COMPILER=rc
+      run: |
+        cmake . -B build %CMAKE_ARGS%
+        cmake --build build -j %NUMBER_OF_PROCESSORS%

.github/actions/setup-windows/action.yml

@@ -0,0 +1,37 @@
+name: 'Setup Windows environment'
+
+inputs:
+  python-version:
+    description: 'Version of python to set up'
+    required: false
+    default: '3.14'
+  use-ccache:
+    description: 'Whether to enable ccache'
+    required: false
+    default: 'true'
+
+runs:
+  using: 'composite'
+  steps:
+    - name: Use ccache
+      if: ${{ inputs.use-ccache == 'true' }}
+      uses: hendrikmuhs/[email protected]
+      with:
+        key: ccache-${{ runner.os }}-${{ runner.arch }}-cpu
+        max-size: 1GB
+
+    - uses: actions/setup-python@v6
+      with:
+        python-version: ${{ inputs.python-version }}
+
+    - name: Setup Visual Studio cmd
+      shell: cmd
+      run: |
+        :: Find out path to VS.
+        pushd "C:\Program Files (x86)\Microsoft Visual Studio\Installer\"
+        for /f "delims=" %%x in ('.\vswhere.exe -latest -property InstallationPath') do set VSPATH=%%x
+        popd
+        :: Import VS vars.
+        call "%VSPATH%\VC\Auxiliary\Build\vcvarsall.bat" x64
+        :: Export to all steps.
+        >>%GITHUB_ENV% set

.github/actions/test-windows/action.yml

@@ -0,0 +1,13 @@
+name: 'Run tests on Windows'
+
+runs:
+  using: 'composite'
+  steps:
+    - name: Run CPP tests - CPU
+      shell: bash
+      env:
+        DEVICE: cpu
+      run: |
+        echo "::group::CPP tests - CPU"
+        ./build/tests/tests -tce="*gguf*,test random uniform"
+        echo "::endgroup::"

.github/workflows/build_and_test.yml

@@ -76,6 +76,18 @@ jobs:
       - uses: ./.github/actions/setup-macos
       - uses: ./.github/actions/build-macos
 
+  windows_build_and_test:
+    name: Windows (cpu, x86_64)
+    needs: check_lint
+    runs-on: windows-2025
+    steps:
+      - uses: actions/checkout@v6
+      - uses: ./.github/actions/setup-windows
+        with:
+          use-ccache: false
+      - uses: ./.github/actions/build-windows
+      - uses: ./.github/actions/test-windows
+
   build_documentation:
     name: Build Documentation
     if: github.repository == 'ml-explore/mlx'

CMakeLists.txt

@@ -150,10 +150,6 @@ cmake_policy(SET CMP0135 NEW)
 
 add_library(mlx)
 
-# Supress warnings: note: parameter passing for argument of type
-# ‘std::pair<float, float>’ when C++17 is enabled changed to match C++14 in GCC
-# 10.1
-target_compile_options(mlx PRIVATE -Wno-psabi)
 target_compile_options(mlx PUBLIC ${SANITIZER_COMPILE_FLAGS})
 target_link_options(mlx PUBLIC ${SANITIZER_LINK_FLAGS})
 

mlx/CMakeLists.txt

@@ -24,9 +24,21 @@ add_library(mlx_version OBJECT ${CMAKE_CURRENT_SOURCE_DIR}/version.cpp)
 target_compile_definitions(mlx_version PRIVATE MLX_VERSION="${MLX_VERSION}")
 target_link_libraries(mlx PRIVATE $<BUILD_INTERFACE:mlx_version>)
 
+if(CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
+  # Supress warnings: note: parameter passing for argument of type
+  # ‘std::pair<float, float>’ when C++17 is enabled changed to match C++14 in
+  # GCC 10.1
+  target_compile_options(mlx PRIVATE -Wno-psabi)
+endif()
+
 if(MSVC)
   # Disable some MSVC warnings to speed up compilation.
   target_compile_options(mlx PUBLIC /wd4068 /wd4244 /wd4267 /wd4804)
+  # Enable /bigobj for heavily templated code (e.g., binary.cpp) that exceeds
+  # the default 65,535 section limit in COFF object files. Use generator
+  # expression to only apply to C/CXX, not CUDA (NVCC doesn't understand /bigobj
+  # directly).
+  target_compile_options(mlx PRIVATE $<$<COMPILE_LANGUAGE:CXX>:/bigobj>)
 endif()
 
 if(WIN32)

mlx/backend/cpu/indexing.cpp

@@ -761,7 +761,7 @@ void masked_scatter_impl(const array& mask, const array& src, array& out) {
   const size_t mask_batch_size = mask.size() / batch_count;
   const size_t src_batch_size = src.size() / batch_count;
 
-  for (uint b = 0; b < batch_count; ++b) {
+  for (size_t b = 0; b < batch_count; ++b) {
     size_t src_consumed = 0;
     src_it.seek(b * src_batch_size);
 

mlx/backend/cpu/simd/base_simd.h

@@ -6,6 +6,10 @@
 #include <complex>
 #include <functional>
 
+#ifdef _MSC_VER
+#include <intrin.h> // For _BitScanReverse
+#endif
+
 namespace mlx::core::simd {
 template <typename T, int N>
 struct Simd;
@@ -105,7 +109,7 @@ Simd<T, 1> log1p(Simd<T, 1> in) {
       if (r == 0) { // handle underflow
         return Simd<T, 1>{T{x, theta}};
       }
-      return Simd<T, 1>{T{((typeof(x))(0.5)) * std::log1p(r), theta}};
+      return Simd<T, 1>{T{((decltype(x))(0.5)) * std::log1p(r), theta}};
     } else {
       auto z0 = std::hypot(x + 1, y);
       return Simd<T, 1>{T{std::log(z0), theta}};
@@ -173,7 +177,16 @@ DEFAULT_BINARY(||)
 
 template <typename T>
 Simd<T, 1> clz(Simd<T, 1> x_) {
+#ifdef _MSC_VER
+  // MSVC doesn't have __builtin_clz, use _BitScanReverse instead
+  unsigned long index;
+  if (_BitScanReverse(&index, static_cast<unsigned long>(x_.value))) {
+    return static_cast<T>(31 - index);
+  }
+  return static_cast<T>(32); // All zeros case
+#else
   return __builtin_clz(x_.value);
+#endif
 }
 
 template <typename T>