build llama-cpp on CI, windows + arm64

.github/workflows/ci.yml

@@ -73,6 +73,12 @@ jobs:
           conan install . --build=missing -cc core.version_ranges:resolve_prereleases=True
           conan build . -cc core.version_ranges:resolve_prereleases=True
 
+      # Build llama-cpp for Windows arm64, both natively and crossbuilt from x86_64
+      - name: conan create llama-cpp
+        if: matrix.cuda-version == '13.4' && startsWith(matrix.os, 'windows')
+        working-directory: samples/llama-cpp
+        run: conan create all --version=b6565 -cc core.version_ranges:resolve_prereleases=True --build=missing -pr clang-cl-arm64
+
       # Skip the following for 13.4-preview
       - name: conan create cudnn
         if: matrix.cuda-version != '13.4' && !contains(matrix.os, '-arm')

docs/building-for-windows-on-arm.md

@@ -63,7 +63,7 @@ conan create . -s arch=armv8
 
 When cross-building, Conan ensures that `nvcc` runs on your x86_64 workstation, while the CUDA libraries linked into your application are ARM64.
 
-## Example: building and running `cuda-samples`
+## Example 1: building and running `cuda-samples`
 
 [`cuda-samples`](https://github.com/NVIDIA/cuda-samples) is NVIDIA's official collection of example CUDA
 applications. It ranges from minimal examples like
@@ -103,6 +103,27 @@ build\Release\cpp\1_Utilities\deviceQuery\deviceQuery.exe
 The application should launch successfuly. If there is no GPU with an NVIDIA driver installed, it should simply report that cudaGetDeviceCount returned an error querying the devices.
 
 
+## Example 2: building `llama-cpp` for Windows ARM64 with CUDA support for RTX Spark
+
+[`llama.cpp`](https://github.com/ggml-org/llama.cpp) is a widely used, dependency-light C/C++ inference engine
+for LLaMA and other GGUF-format language models. It's one of the most popular ways to run LLMs locally, with
+over 80k stars on GitHub, and is the engine behind many downstream projects (Ollama, LM Studio, and others).
+
+This repository contains a recipe to build `llama.cpp` from source using CUDA from Conan, including support
+for Windows ARM64:
+
+```
+cd samples/llama-cpp
+conan create all --version=b6565 -pr clang-cl-arm64 -cc core.version_ranges:resolve_prereleases=True
+```
+
+This command can be run on x86_64 Windows and ARM64 Windows. You can customize the CUDA GPU architecture by editing the profile at `samples/llama-cpp/clang-cl-common`.
+
+> [!IMPORTANT]
+> Windows ARM64 builds of `llama.cpp` currently currently require building with **Clang** (`clang-cl`), rather than MSVC's
+> `cl.exe`. Clang can be installed via the **Visual Studio Installer**: under **Individual components**, search
+> for **"Clang"** and select the **C++ Clang Compiler for Windows** component, then apply.
+
 ## Current limitations
 
 - While the ARM64 binaries will run on any Windows on ARM, CUDA support at runtime can only be tested on NVIDIA-powered devices.

samples/llama-cpp/all/conanfile.py

@@ -91,6 +91,7 @@ def generate(self):
         # Follow with_examples when newer versions can compile examples,
         # right now it tries to add_subdirectory to a non-existent folder
         tc.variables["GGML_BUILD_EXAMPLES"] = False
+        tc.cache_variables["GGML_OPENMP"] = "OFF"
         tc.variables["GGML_CUDA"] = self.options.get_safe("with_cuda")
         if self.settings.compiler in ("msvc", "clang") and self.options.get_safe("with_cuda"):
             tc.cache_variables['CMAKE_CUDA_FLAGS'] = '-Xcompiler /Zc:preprocessor '

samples/llama-cpp/clang-cl-arm64

@@ -0,0 +1,8 @@
+include(clang-cl-common)
+
+[settings]
+arch=armv8
+
+[conf]
+tools.build:cflags=['--target=aarch64-pc-windows-msvc']
+tools.build:cxxflags=['--target=aarch64-pc-windows-msvc']

samples/llama-cpp/clang-cl-common

@@ -0,0 +1,33 @@
+{% set msvc, msvc_version, _ = detect_api.detect_msvc_compiler() %}
+{# msvc_version is e.g. "193", "194", "195" — first two digits = toolset generation #}
+{% set toolset_map = {"190": "v140", "191": "v141", "192": "v142", "193": "v143", "194": "v144", "195": "v145"} %}
+{% set runtime_version = toolset_map.get(msvc_version|string|truncate(3, end=''), "v143") %}
+
+{# Derive the VS install path from the registry via vswhere #}
+{% set vswhere = subprocess.check_output(
+    ['C:\\Program Files (x86)\\Microsoft Visual Studio\\Installer\\vswhere.exe',
+     '-latest', '-property', 'installationPath'],
+    stderr=subprocess.DEVNULL
+).decode().strip() %}
+
+{# Read the clang-cl version from its manifest file inside the VS install — no dev prompt needed #}
+{% set clang_exe = vswhere + '\\VC\\Tools\\Llvm\\x64\\bin\\clang-cl.exe' %}
+{% set clang_version = subprocess.check_output(
+    [clang_exe, '--version'],
+    stderr=subprocess.STDOUT
+).decode().split('version ')[1].split('.')[0].strip() %}
+
+[settings]
+build_type=Release
+compiler=clang
+compiler.cppstd=17
+compiler.runtime=dynamic
+compiler.version={{ clang_version }}
+compiler.runtime_version={{ runtime_version }}
+os=Windows
+
+
+[conf]
+tools.cmake.cmaketoolchain:generator=Ninja
+tools.build:compiler_executables={"c": "clang-cl", "cpp": "clang-cl"}
+tools.cmake.cmaketoolchain:extra_variables*={'CMAKE_CUDA_ARCHITECTURES':'87-real'}

samples/llama-cpp/clang-cl-x64

@@ -0,0 +1,4 @@
+include(clang-cl-common)
+
+[settings]
+arch=x86_64