feat(windows-cuda): make CUDA for Windows

Daniele Briggi · Daniele Briggi · commit 75fa13ab7ea4 · 2025-07-18T12:23:25.000+02:00
diff --git a/.github/workflows/main.yml b/.github/workflows/main.yml
@@ -15,30 +15,34 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - name: linux-vulkan
-            os: ubuntu-latest
-            arch: x86_64
-            make: LLAMA="-DGGML_BACKEND_DL=ON -DGGML_NATIVE=OFF -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON"
-          - name: macos
-            os: macos-latest
-          - name: windows
+          # - name: linux-vulkan
+          #   os: ubuntu-latest
+          #   arch: x86_64
+          #   make: LLAMA="-DGGML_BACKEND_DL=ON -DGGML_NATIVE=OFF -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON"
+          # - name: macos
+          #   os: macos-latest
+          # - name: windows
+          #   os: windows-latest
+          #   arch: x86_64
+          - name: windows-cuda
             os: windows-latest
             arch: x86_64
-          - name: android
-            os: ubuntu-latest
-            arch: arm64-v8a
-            make: PLATFORM=android ARCH=arm64-v8a
-          - name: android
-            os: ubuntu-latest
-            arch: x86_64
-            make: PLATFORM=android ARCH=x86_64
-            sqlite-amalgamation-zip: https://sqlite.org/2025/sqlite-amalgamation-3490100.zip
-          - name: ios
-            os: macos-latest
-            make: PLATFORM=ios
-          - name: isim
-            os: macos-latest
-            make: PLATFORM=isim
+            make: LLAMA="-DGGML_BACKEND_DL=ON -DGGML_NATIVE=OFF -DGGML_CPU=OFF -DGGML_CUDA=ON"
+          # - name: android
+          #   os: ubuntu-latest
+          #   arch: arm64-v8a
+          #   make: PLATFORM=android ARCH=arm64-v8a
+          # - name: android
+          #   os: ubuntu-latest
+          #   arch: x86_64
+          #   make: PLATFORM=android ARCH=x86_64
+          #   sqlite-amalgamation-zip: https://sqlite.org/2025/sqlite-amalgamation-3490100.zip
+          # - name: ios
+          #   os: macos-latest
+          #   make: PLATFORM=ios
+          # - name: isim
+          #   os: macos-latest
+          #   make: PLATFORM=isim
 
     defaults:
       run:
@@ -50,15 +54,6 @@ jobs:
         with:
           submodules: true
 
-      - uses: msys2/setup-msys2@v2.27.0
-        if: matrix.os == 'windows-latest'
-        with:
-          msystem: mingw64
-          install: >-
-            mingw-w64-x86_64-cc
-            mingw-w64-x86_64-cmake
-            make
-
       - name: linux install dependencies
         if: matrix.name == 'linux-vulkan'
         run: |
@@ -82,17 +77,20 @@ jobs:
           echo "LD_LIBRARY_PATH=$LD_LIBRARY_PATH" >> $GITHUB_ENV
           echo "VK_ADD_LAYER_PATH=$VK_ADD_LAYER_PATH" >> $GITHUB_ENV
 
-      - name: windows install dependencies
-        if: matrix.os == 'windows-latest'
+
+      - name: Install CUDA Toolkit
+        if: matrix.name == 'windows-cuda'
         run: |
-          pacman -S --noconfirm git sqlite3
-        shell: msys2 {0}
+          choco install cuda -y
+        shell: powershell
 
-      - name: windows build sqlite-ai
-        if: matrix.name == 'windows'
-        run: make extension ${{ matrix.make && matrix.make || ''}}
-        shell: msys2 {0}
-          
+      - name: Windows build sqlite-ai
+        if: matrix.os == 'windows-latest' || matrix.name == 'windows-cuda'
+        run: |
+          call "%ProgramFiles(x86)%\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvars64.bat"
+          nmake /f Makefile PLATFORM=windows ${{ matrix.make && matrix.make || ''}}
+        shell: cmd
+      
       - name: build sqlite-ai
         if: matrix.name != 'windows'
         run: make extension ${{ matrix.make && matrix.make || ''}}
@@ -143,10 +141,13 @@ jobs:
             adb push ${{ github.workspace }}/. /data/local/tmp/
             adb shell "sh /data/local/tmp/commands.sh"
 
+
       - name: windows test sqlite-ai
-        if: matrix.name == 'windows'
-        run: make test
-        shell: msys2 {0}
+        if: matrix.os == 'windows-latest'
+        run: |
+          call "%ProgramFiles(x86)%\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvars64.bat"
+          nmake /f Makefile test PLATFORM=windows
+        shell: cmd
 
       - name: test sqlite-ai
         if: matrix.name == 'linux' || matrix.name == 'macos'
diff --git a/Makefile b/Makefile
@@ -36,6 +36,7 @@ BUILD_LLAMA = $(BUILD_DIR)/llama.cpp
 BUILD_WHISPER = $(BUILD_DIR)/whisper.cpp
 BUILD_MINIAUDIO = $(BUILD_DIR)/miniaudio
 
+OBJ_EXT = o
 # Compiler and flags
 CC = gcc
 CXX = g++
@@ -53,18 +54,21 @@ LDFLAGS = $(LLAMA_LDFLAGS) $(WHISPER_LDFLAGS) $(MINIAUDIO_LDFLAGS)
 
 # Files
 SRC_FILES = $(wildcard $(SRC_DIR)/*.c)
-OBJ_FILES = $(patsubst %.c, $(BUILD_DIR)/%.o, $(notdir $(SRC_FILES)))
+OBJ_FILES = $(patsubst %.c, $(BUILD_DIR)/%.$(OBJ_EXT), $(notdir $(SRC_FILES)))
 LLAMA_LIBS = $(BUILD_LLAMA)/common/libcommon.a $(BUILD_LLAMA)/ggml/src/libggml.a $(BUILD_LLAMA)/ggml/src/libggml-base.a $(BUILD_LLAMA)/ggml/src/libggml-cpu.a $(BUILD_LLAMA)/src/libllama.a
 WHISPER_LIBS = $(BUILD_WHISPER)/src/libwhisper.a
 MINIAUDIO_LIBS = $(BUILD_MINIAUDIO)/libminiaudio.a
 
 # Platform-specific settings
 ifeq ($(PLATFORM),windows)
+	OBJ_EXT = obj
+	CC = cl
+	CXX = cl
 	TARGET := $(DIST_DIR)/ai.dll
-	LDFLAGS = -L./$(BUILD_LLAMA)/common -L./$(BUILD_LLAMA)/ggml/src -L./$(BUILD_LLAMA)/src -L./$(BUILD_WHISPER)/src -L./$(BUILD_MINIAUDIO) -l:libllama.a -l:libwhisper.a -l:libminiaudio.a -l:libcommon.a -l:ggml.a -l:ggml-cpu.a -l:ggml-base.a -fopenmp -static-libgcc -static-libstdc++ -shared -lbcrypt
-	# Create .def file for Windows
+	CFLAGS = /nologo /W3 /EHsc /I$(SRC_DIR) /I$(LLAMA_DIR)/ggml/include /I$(LLAMA_DIR)/include /I$(WHISPER_DIR)/include /I$(MINIAUDIO_DIR) /MD
+	LDFLAGS = /DLL /OUT:$(TARGET) /LIBPATH:$(BUILD_LLAMA)/common /LIBPATH:$(BUILD_LLAMA)/ggml/src /LIBPATH:$(BUILD_LLAMA)/src /LIBPATH:$(BUILD_WHISPER)/src /LIBPATH:$(BUILD_MINIAUDIO) common.lib ggml.lib ggml-cpu.lib ggml-base.lib llama.lib whisper.lib miniaudio.lib bcrypt.lib
 	DEF_FILE := $(BUILD_DIR)/ai.def
-	STRIP = strip --strip-unneeded $@
+	STRIP = echo "No strip needed for MSVC"
 else ifeq ($(PLATFORM),macos)
 	TARGET := $(DIST_DIR)/ai.dylib
 	LLAMA_LIBS += $(BUILD_LLAMA)/ggml/src/ggml-metal/libggml-metal.a $(BUILD_LLAMA)/ggml/src/ggml-blas/libggml-blas.a
@@ -128,7 +132,7 @@ else ifeq ($(PLATFORM),isim)
 	WHISPER_OPTIONS += -DBUILD_SHARED_LIBS=OFF -DCMAKE_SYSTEM_NAME=iOS -DCMAKE_OSX_SYSROOT=iphonesimulator -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 -DCMAKE_OSX_ARCHITECTURES="x86_64;arm64" -DWHISPER_COREML=ON
 else # linux
 	TARGET := $(DIST_DIR)/ai.so
-	# using -DGGML_CPU_ALL_VARIANTS=ON
+	# Using -DGGML_CPU_ALL_VARIANTS=ON, `-lggml-cpu` is not needed
 	LDFLAGS := $(filter-out -lggml-cpu,$(LDFLAGS))
 	LDFLAGS += -shared -L./$(BUILD_LLAMA)/bin -Wl,-rpath,./$(BUILD_LLAMA)/bin -Wl,-rpath,./$(BUILD_LLAMA)/common -Wl,-rpath,./$(BUILD_LLAMA)/ggml/src -Wl,-rpath,./$(BUILD_LLAMA)/src -Wl,-rpath,./$(BUILD_WHISPER)/src
 	# Add miniaudio Linux-specific flags (as per miniaudio docs)
@@ -158,17 +162,21 @@ compile: $(OBJ_FILES)
 
 # Loadable library
 $(TARGET): $(DEF_FILE) $(LLAMA_LIBS) $(WHISPER_LIBS) $(MINIAUDIO_LIBS)
-	$(CXX) $(OBJ_FILES) $(DEF_FILE) -o $@ $(LDFLAGS)
 ifeq ($(PLATFORM),windows)
-	# Generate import library for Windows
-	dlltool -D $@ -d $(DEF_FILE) -l $(DIST_DIR)/ai.lib
+	$(CXX) $(OBJ_FILES) /link $(LDFLAGS)
+else
+	$(CXX) $(OBJ_FILES) $(DEF_FILE) -o $@ $(LDFLAGS)
 endif
-	# Strip debug symbols
 	$(STRIP)
 
 # Object files
+ifeq ($(PLATFORM),windows)
+$(BUILD_DIR)/%.obj: %.c
+	$(CC) $(CFLAGS) /c $< /Fo$@
+else
 $(BUILD_DIR)/%.o: %.c
 	$(CC) $(CFLAGS) -O3 -fPIC -c $< -o $@
+endif
 
 test: $(TARGET)
 	$(SQLITE3) ":memory:" -cmd ".bail on" ".load ./dist/ai" "SELECT ai_version();"
