feat(windows-cuda): make CUDA for Windows

Author: Daniele Briggi

.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:
@@ -89,7 +93,7 @@ jobs:
         shell: msys2 {0}
 
       - name: windows build sqlite-ai
-        if: matrix.name == 'windows'
+        if: matrix.name == 'windows' || matrix.name == 'windows-cuda'
         run: make extension ${{ matrix.make && matrix.make || ''}}
         shell: msys2 {0}
 

Makefile

@@ -128,7 +128,7 @@ 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)