ci: add Windows CUDA build

Author: KakaruHayate

.github/workflows/ci.yml

@@ -8,6 +8,7 @@ name: ci
 #   macos-arm64-metal  — Apple Silicon, Metal backend
 #   macos-x64-metal    — Intel Mac, Metal backend (cross-compiled on ARM runner)
 #   windows-x64-vulkan — Windows, Vulkan backend
+#   windows-x64-cuda   — Windows Server 2022, CUDA backend (CUDA 12.6)
 #
 # On tag push, artifacts are bundled into a GitHub Release.
 
@@ -97,34 +98,47 @@ jobs:
             name: linux-x64-vulkan
             backend: vulkan
             cmake_extra: "-DGAME_GGML_VULKAN=ON"
+            build_jobs: 4
             pkg_ext: ""
             lib_glob: "libggml*.so*"
 
           - os: ubuntu-22.04
             name: linux-x64-cuda
             backend: cuda
-            cmake_extra: '-DGAME_GGML_CUDA=ON -DGGML_NATIVE=OFF -DCMAKE_CUDA_ARCHITECTURES=75\;80\;86\;89'
+            cmake_extra: '-DGAME_GGML_CUDA=ON -DGGML_NATIVE=OFF -DCMAKE_CUDA_ARCHITECTURES=75-real\;80-real\;86-real\;89-real'
+            build_jobs: 2
             pkg_ext: ""
             lib_glob: "libggml*.so*"
 
           - os: macos-14
             name: macos-arm64-metal
             backend: metal
             cmake_extra: "-DGAME_GGML_METAL=ON -DCMAKE_OSX_ARCHITECTURES=arm64"
+            build_jobs: 4
             pkg_ext: ""
             lib_glob: "libggml*.dylib"
 
           - os: macos-14
             name: macos-x64-metal
             backend: metal
             cmake_extra: "-DGAME_GGML_METAL=ON -DCMAKE_OSX_ARCHITECTURES=x86_64"
+            build_jobs: 4
             pkg_ext: ""
             lib_glob: "libggml*.dylib"
 
           - os: windows-latest
             name: windows-x64-vulkan
             backend: vulkan
             cmake_extra: "-DGAME_GGML_VULKAN=ON"
+            build_jobs: 4
+            pkg_ext: ".exe"
+            lib_glob: "ggml*.dll"
+
+          - os: windows-2022
+            name: windows-x64-cuda
+            backend: cuda
+            cmake_extra: '-DGAME_GGML_CUDA=ON -DGGML_NATIVE=OFF -DCMAKE_CUDA_ARCHITECTURES=75-real\;80-real\;86-real\;89-real'
+            build_jobs: 2
             pkg_ext: ".exe"
             lib_glob: "ggml*.dll"
 
@@ -147,12 +161,21 @@ jobs:
         if: matrix.backend == 'vulkan'
         run: glslc --version
 
-      - name: Install CUDA Toolkit
-        if: matrix.backend == 'cuda'
-        uses: Jimver/cuda-toolkit@1a3c14e26833ccf292b268f9a790fb47dea7b2da
+      - name: Install CUDA Toolkit (Linux)
+        if: matrix.backend == 'cuda' && runner.os == 'Linux'
+        uses: Jimver/cuda-toolkit@1a3c14e26833ccf292b268f9a790fb47dea7b2da # v0.2.28
+        with:
+          cuda: "12.6.3"
+          method: network
+          log-file-suffix: "${{ matrix.name }}.txt"
+
+      - name: Install CUDA Toolkit (Windows)
+        if: matrix.backend == 'cuda' && runner.os == 'Windows'
+        uses: Jimver/cuda-toolkit@b8bf9c6c28f8a92fbb04dcfcaee872e60c57462d # v0.2.36
         with:
           cuda: "12.6.3"
           method: network
+          sub-packages: '["nvcc", "cudart", "cublas", "cublas_dev", "visual_studio_integration"]'
           log-file-suffix: "${{ matrix.name }}.txt"
 
       - name: Verify CUDA compiler
@@ -181,7 +204,9 @@ jobs:
                 ${{ matrix.cmake_extra }}
 
       - name: Build
-        run: cmake --build build -j --config Release
+        # ggml-cuda expands many template instances. Unbounded parallel builds can
+        # exhaust the memory of GitHub-hosted runners and terminate nvcc.
+        run: cmake --build build --parallel ${{ matrix.build_jobs }} --config Release
 
       - name: Verify binary
         run: |

BUILDING.md

@@ -20,7 +20,8 @@ sudo apt install libvulkan-dev vulkan-tools
 #   sudo apt install glslc-tools   # not available on all distros
 
 # Optional: CUDA backend (NVIDIA GPUs)
-# Install a CUDA Toolkit supported by your compiler and driver. CI uses CUDA 12.6.
+# Install a CUDA Toolkit supported by your compiler and driver. CI uses CUDA 12.6.3.
+# The prebuilt CUDA packages target Turing (CC 7.5) and newer GPUs.
 # https://developer.nvidia.com/cuda-downloads
 
 # Optional: ccache for faster rebuilds
@@ -53,6 +54,10 @@ brew install cmake ccache
 
 # Vulkan SDK (optional, for Vulkan backend)
 #   https://vulkan.lunarg.com/sdk/home
+#
+# CUDA Toolkit 12.6.x (optional, for CUDA backend)
+#   https://developer.nvidia.com/cuda-downloads
+# CUDA 12.6 supports Visual Studio 2022 / MSVC 193x.
 ```
 
 ## Quick start
@@ -100,7 +105,7 @@ cmake -B build -DCMAKE_BUILD_TYPE=Release \
                -DGAME_GGML_BUILD_CLI=ON \
                -DGAME_GGML_CUDA=ON \
                -DGGML_NATIVE=OFF \
-               -DCMAKE_CUDA_ARCHITECTURES="75;80;86;89"
+               -DCMAKE_CUDA_ARCHITECTURES="75-real;80-real;86-real;89-real"
 
 # macOS Apple Silicon + Metal
 cmake -B build -DCMAKE_BUILD_TYPE=Release \
@@ -117,6 +122,13 @@ cmake -B build -DCMAKE_BUILD_TYPE=Release \
 cmake -B build -DCMAKE_BUILD_TYPE=Release `
                -DGAME_GGML_BUILD_CLI=ON `
                -DGAME_GGML_VULKAN=ON
+
+# Windows + CUDA (from Visual Studio 2022 Developer PowerShell)
+cmake -B build -DCMAKE_BUILD_TYPE=Release `
+               -DGAME_GGML_BUILD_CLI=ON `
+               -DGAME_GGML_CUDA=ON `
+               -DGGML_NATIVE=OFF `
+               -DCMAKE_CUDA_ARCHITECTURES="75-real;80-real;86-real;89-real"
 ```
 
 ## Converting a PyTorch checkpoint to GGUF
@@ -163,14 +175,33 @@ build/bin/game_ggml_cli serve game_medium.gguf
 # Then write binary request frames to stdin (see src/cli/main.cpp for protocol)
 ```
 
-## CI CUDA scope
-
-The hosted CI builds and packages the Linux x64 CUDA backend with CUDA Toolkit
-12.6. It verifies Toolkit discovery, CUDA compilation, linking, and that the CLI
-starts with `--version`. GitHub-hosted runners do not provide an NVIDIA GPU, so
-actual CUDA inference must still be smoke-tested on an NVIDIA system. The
-packaged CUDA backend uses the CUDA runtime and cuBLAS libraries supplied by the
-installed NVIDIA CUDA runtime/toolkit.
+## CUDA compatibility and CI scope
+
+The hosted CI builds Linux x64 and Windows x64 CUDA packages with CUDA Toolkit
+12.6.3 and Visual Studio 2022 on Windows. It verifies Toolkit discovery, CUDA
+compilation, linking, and that the CLI starts with `--version`. GitHub-hosted
+runners do not provide an NVIDIA GPU, so actual CUDA inference must still be
+smoke-tested on an NVIDIA system.
+
+The release architecture list is `75-real;80-real;86-real;89-real`, covering:
+
+- CC 7.5: Turing (for example, GeForce RTX 20 series)
+- CC 8.0/8.6: Ampere (A100 and GeForce RTX 30 series)
+- CC 8.9: Ada (GeForce RTX 40 series)
+
+`-real` emits native SASS for each target and avoids requiring the display driver
+to JIT PTX generated by CUDA 12.6. Pascal and Volta are not included in the
+prebuilt package. Source builds that need these older GPUs can use CUDA 12.x and
+add `61-real` and/or `70-real`. CUDA 13.0 removed NVCC offline compilation for
+architectures older than CC 7.5; use CUDA 12.9 or earlier when maintaining such
+builds.
+
+CUDA 12.x minor-version compatibility requires at least NVIDIA driver
+525.60.13 on Linux or 528.33 on Windows, subject to the limitations documented
+in NVIDIA's CUDA Compatibility Guide. Using the current production driver is
+recommended. The packages currently expect the CUDA 12 runtime and cuBLAS
+libraries to be installed on the target system; they do not bundle NVIDIA's
+runtime libraries.
 
 ## Troubleshooting