Build for Android (Experimental)

Step-by-step instructions for cross-compiling llama.cpp for Android (ARM64) for use with GenAI Llama’s embedded inference.

Prerequisites

Android NDK (r25 or later recommended)
- Included with Android Studio, or download standalone from developer.android.com/ndk
CMake 3.14 or later
Git

Set the ANDROID_NDK environment variable to point to your NDK installation:

export ANDROID_NDK=/path/to/android-ndk-r25c

Choose Your GPU Backend

Backend	Best For	Flag
Vulkan	Most modern Android devices (recommended)	`-DGGML_VULKAN=ON`
CPU	Older devices, or testing	(no GPU flag)

Vulkan Support

Most Android devices released since 2017 support Vulkan. For maximum compatibility with older devices, build with CPU-only as a fallback.

Build Steps

1. Clone llama.cpp

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
git checkout b3600  # or any later release tag

2. Configure and Build

Vulkan (Recommended)

cmake -B build \
  -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake \
  -DANDROID_ABI=arm64-v8a \
  -DANDROID_PLATFORM=android-26 \
  -DBUILD_SHARED_LIBS=OFF \
  -DGGML_VULKAN=ON
cmake --build build --config Release

CPU Only

cmake -B build \
  -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake \
  -DANDROID_ABI=arm64-v8a \
  -DANDROID_PLATFORM=android-26 \
  -DBUILD_SHARED_LIBS=OFF
cmake --build build --config Release

Important

Always use arm64-v8a as the ABI. This matches the ARM64 architecture required by modern Android devices and Unreal Engine's Android target.

3. Copy Libraries

Copy all .a files from build/ (and its subdirectories) into your plugin directory:

<YourProject>/Plugins/GenAILlama/ThirdParty/LlamaCpp/lib/Android/

find build -name "*.a" -exec cp {} /path/to/ThirdParty/LlamaCpp/lib/Android/ \;

4. Rebuild Your Unreal Project

After placing the files, rebuild your Unreal project targeting Android. Check the build log for:

GenAILlama: llama.cpp libraries found at <path>. Embedded inference enabled.

Model Size Considerations

Mobile devices have limited RAM and no swap. Choose models carefully:

Model	Q4_K_M Size	RAM Usage (Approx.)	Recommended For
Qwen2.5-0.5B	~400 MB	~600 MB	Low-end devices
TinyLlama 1.1B	~637 MB	~900 MB	Mid-range devices
Gemma-2B	~1.4 GB	~2 GB	High-end devices

Set GPU Layers = 0 (CPU only) if you encounter GPU memory issues on specific devices.

Troubleshooting

CMake can’t find the NDK toolchain

Verify ANDROID_NDK points to the correct directory.
The toolchain file should be at $ANDROID_NDK/build/cmake/android.toolchain.cmake.

Vulkan build fails

Ensure your NDK version includes Vulkan headers (NDK r21+ includes them).
Try building with CPU-only first to isolate the issue.

App crashes on older devices

Check that the device supports the minimum Android API level (26).
Try reducing GPU Layers or using CPU-only inference.

Muddy Terrain Games