Add Windows ARM64 wheels

.github/workflows/ci.yml

@@ -189,6 +189,9 @@ jobs:
         include:
         - os: ubuntu-22.04-arm
           arch: aarch64
+        - os: windows-11-arm
+          arch: ARM64
+          build: "cp31[2-4]-*"
         - os: macos-15
           arch: arm64
         - os: macos-15-intel
@@ -200,7 +203,7 @@ jobs:
           submodules: recursive
 
       - name: Cache MKL and oneDNN (Windows)
-        if: runner.os == 'Windows'
+        if: runner.os == 'Windows' && matrix.arch != 'ARM64'
         uses: actions/cache@v4
         with:
           path: |
@@ -224,6 +227,7 @@ jobs:
           CIBW_MANYLINUX_X86_64_IMAGE: manylinux_2_28
           CIBW_MANYLINUX_AARCH64_IMAGE: manylinux_2_28
           CIBW_ARCHS: ${{ matrix.arch }}
+          CIBW_BUILD: ${{ matrix.build || '*' }}
           CIBW_SKIP: "*-musllinux_*"
 
       - name: Upload Python wheels
@@ -295,6 +299,11 @@ jobs:
             artifact_pattern: python-wheels-Windows-auto64
             wheel_pattern: "*cp312*win*amd64.whl"
 
+          - os: windows-11-arm
+            artifact_pattern: python-wheels-Windows-ARM64
+            wheel_pattern: "*cp312*win*arm64.whl"
+            python_arch: arm64
+
           - os: macos-15
             artifact_pattern: python-wheels-macOS-arm64
             wheel_pattern: "*cp312*macosx*arm64.whl"
@@ -304,6 +313,7 @@ jobs:
         uses: actions/setup-python@v5
         with:
           python-version: "3.12"
+          architecture: ${{ matrix.python_arch }}
 
       - uses: actions/checkout@v4
 
@@ -328,6 +338,36 @@ jobs:
         run: |
           pytest -v python/tests/ --ignore=python/tests/test_fairseq.py --ignore=python/tests/test_opennmt_py.py --ignore=python/tests/test_opennmt_tf.py
 
+      - name: Test multithreaded INT8 inference (Windows ARM64)
+        if: runner.os == 'Windows' && runner.arch == 'ARM64'
+        shell: pwsh
+        run: |
+          python -c "import ctranslate2; model = 'tests/data/models/v2/aren-transliteration-i8'; translator = ctranslate2.Translator(model, device='cpu', compute_type='int8', inter_threads=1, intra_threads=4); result = translator.translate_batch([['آ', 'ت', 'ز', 'م', 'و', 'ن']], beam_size=1); assert result[0].hypotheses[0] == ['a', 't', 'z', 'm', 'o', 'n']; del result, translator"
+
+      - name: Test multithreaded Ruy context shutdown (Windows ARM64)
+        if: runner.os == 'Windows' && runner.arch == 'ARM64'
+        timeout-minutes: 2
+        shell: pwsh
+        run: |
+          @'
+          import ctranslate2
+
+          model = "tests/data/models/v2/aren-transliteration"
+          for _ in range(3):
+              translator = ctranslate2.Translator(
+                  model,
+                  device="cpu",
+                  compute_type="float32",
+                  inter_threads=1,
+                  intra_threads=4,
+              )
+              result = translator.translate_batch(
+                  [["آ", "ت", "ز", "م", "و", "ن"]], beam_size=1
+              )
+              assert result[0].hypotheses[0] == ["a", "t", "z", "m", "o", "n"]
+              del result, translator
+          '@ | python -
+
 
   test-legacy-converters:
     needs: [build-python-wheels]

CMakeLists.txt

@@ -272,7 +272,7 @@ macro(ct2_compile_kernels_for_isa isa flag)
   list(APPEND SOURCES ${CMAKE_CURRENT_BINARY_DIR}/kernels_${isa}.cc)
 endmacro()
 
-if(CMAKE_SYSTEM_PROCESSOR MATCHES "(arm64)|(aarch64)"
+if(CMAKE_SYSTEM_PROCESSOR MATCHES "(arm64)|(aarch64)|(ARM64)"
     OR (APPLE AND CMAKE_OSX_ARCHITECTURES STREQUAL "arm64"))
   add_definitions(-DCT2_ARM64_BUILD)
   set(CT2_BUILD_ARCH "arm64")

docs/installation.md

@@ -10,18 +10,18 @@ pip install ctranslate2
 
 The Python wheels have the following requirements:
 
-* OS: Linux (x86-64, AArch64), macOS (x86-64, ARM64), Windows (x86-64)
-* Python version: >= 3.9
+* OS: Linux (x86-64, AArch64), macOS (x86-64, ARM64), Windows (x86-64, ARM64)
+* Python version: >= 3.9 (3.12–3.14 for Windows ARM64)
 * pip version: >= 19.3 to support `manylinux2014` wheels
 
 ```{admonition} GPU support
-The Linux and Windows Python wheels support GPU execution. Install [CUDA](https://developer.nvidia.com/cuda-toolkit) 12.x to use the GPU.
+The Linux and Windows x86-64 Python wheels support GPU execution. Windows ARM64 wheels are CPU-only. Install [CUDA](https://developer.nvidia.com/cuda-toolkit) 12.x to use the GPU.
 
 If you plan to run models with convolutional layers (e.g. for speech recognition), you should also install [cuDNN 8](https://developer.nvidia.com/cudnn) for CUDA 12.x.
 ```
 
 ```{note}
-On Windows [the Visual C++ runtime](https://www.microsoft.com/en-US/download/details.aspx?id=48145) is required. It is installed in most of the systems, but if is not the case, you have to download it and install it.
+On Windows [the latest supported Visual C++ runtime](https://learn.microsoft.com/en-us/cpp/windows/latest-supported-vc-redist) is required. It is installed in most systems, but if this is not the case, you have to download and install it.
 ```
 
 ## Install with Docker

python/tools/prepare_build_environment_windows.sh

@@ -3,6 +3,21 @@
 set -e
 set -x
 
+NPROC=$(nproc)
+
+if [ "$CIBW_ARCHS" = "ARM64" ]; then
+    mkdir build
+    cd build
+    cmake -A ARM64 -DCMAKE_POLICY_VERSION_MINIMUM=3.5 -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX="$CTRANSLATE2_ROOT" -DBUILD_CLI=OFF -DWITH_MKL=OFF -DWITH_DNNL=OFF -DWITH_RUY=ON -DWITH_CUDA=OFF -DOPENMP_RUNTIME=COMP ..
+    cmake --build . --config Release --target install --parallel "$NPROC" --verbose
+    cd ..
+    rm -r build
+
+    cp README.md python/
+    cp "$CTRANSLATE2_ROOT/bin/ctranslate2.dll" python/ctranslate2/
+    exit 0
+fi
+
 CUDA_ROOT="C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v12.8"
 curl --netrc-optional -L -nv -o cuda.exe https://developer.download.nvidia.com/compute/cuda/12.8.1/local_installers/cuda_12.8.1_572.61_windows.exe
 ./cuda.exe -s nvcc_12.8 cudart_12.8 cublas_dev_12.8 curand_dev_12.8
@@ -38,7 +53,6 @@ if [ ! -d "C:/Program Files (x86)/Intel/oneAPI" ]; then
     ./webimage_extracted/bootstrapper.exe -s --action install --components="intel.oneapi.win.mkl.devel" --eula=accept -p=NEED_VS2017_INTEGRATION=0 -p=NEED_VS2019_INTEGRATION=0 --log-dir=.
 fi
 
-NPROC=$(nproc)
 ONEDNN_VERSION=3.1.1
 if [ ! -d "C:/Program Files (x86)/oneDNN" ]; then
     curl --netrc-optional -L -O https://github.com/oneapi-src/oneDNN/archive/refs/tags/v${ONEDNN_VERSION}.tar.gz

src/cpu/backend.cc

@@ -1,5 +1,7 @@
 #include "backend.h"
 
+#include <memory>
+
 #include "ctranslate2/utils.h"
 #include "cpu_info.h"
 #include "env.h"
@@ -106,9 +108,22 @@ namespace ctranslate2 {
     }
 
 #ifdef CT2_WITH_RUY
+    static std::unique_ptr<ruy::Context>& get_thread_local_ruy_context() {
+      static thread_local std::unique_ptr<ruy::Context> context;
+      return context;
+    }
+
     ruy::Context *get_ruy_context() {
-      static thread_local ruy::Context context;
-      return &context;
+      auto& context = get_thread_local_ruy_context();
+      if (!context)
+        context = std::make_unique<ruy::Context>();
+      return context.get();
+    }
+
+    void destroy_ruy_context() {
+      // The Ruy context owns worker threads. Destroy it before this thread
+      // enters TLS cleanup, where joining these workers can deadlock on Windows.
+      get_thread_local_ruy_context().reset();
     }
 #endif
   }

src/cpu/backend.h

@@ -28,6 +28,7 @@ namespace ctranslate2 {
     bool pack_gemm_weights(ComputeType compute_type);
 #ifdef CT2_WITH_RUY
     ruy::Context *get_ruy_context();
+    void destroy_ruy_context();
 #endif
 
   }

src/cpu/vec_neon.h

@@ -160,7 +160,7 @@ namespace ctranslate2 {
       }
 
       static inline value_type round(value_type v) {
-#ifdef __aarch64__
+#if defined(__aarch64__) || defined(_M_ARM64)
         return vrndiq_f32(v);
 #else
         float temp[4] = {std::nearbyintf(v[0]), std::nearbyintf(v[1]), std::nearbyintf(v[2]), std::nearbyintf(v[3])};

src/devices.cc

@@ -8,6 +8,7 @@
 #  include <unistd.h>
 #endif
 
+#include "cpu/backend.h"
 #include "device_dispatch.h"
 
 namespace ctranslate2 {
@@ -121,6 +122,11 @@ namespace ctranslate2 {
   }
 
   void destroy_context(Device device) {
+#ifdef CT2_WITH_RUY
+      if (device == Device::CPU) {
+        cpu::destroy_ruy_context();
+      }
+#endif
 #ifdef CT2_WITH_CUDA
       if (device == Device::CUDA) {
           cuda::free_curand_states();