Docs

Mobile

Android and iOS use the phone preset through the C API. Success is peak RSS and the process staying up.

Preset

About 2 GiB, one staging buffer, 4 threads, KV window 512 plus 4 sink tokens. The budget is then half of what the OS can still give this process. Details and the laptop rule are on Memory.

GPU

iOS Metal runs MatMul. Android Vulkan runs mapped f32 MatMul, including transpose_b. Attention and RoPE stay on CPU. Q4_K stays in the CPU quant GEMM, which uses NEON on arm64. The sample selects the backend with auto.

C API

uaii_memory_preset_info info;
memset(&info, 0, sizeof info);
info.struct_size = sizeof info;
uaii_resolve_memory_preset("phone", 0, &info);

uaii_session_options opts;
uaii_session_options_init(&opts);
uaii_session_options_apply_preset(&opts, &info);
opts.backend = "auto";

uaii_session_set_background(session, 1) makes generate and run return paused: app backgrounded. uaii_session_request_cancel returns cancelled. Generate clears cancel when it starts. uaii_session_memory_warning drops KV and keeps the longrope scale length. Callers built against 0.3.0 keep working: new fields are appended.

Android

cmake -S examples/mobile -B build-android -G Ninja \
  -DCMAKE_TOOLCHAIN_FILE="$ANDROID_NDK/build/cmake/android.toolchain.cmake" \
  -DANDROID_ABI=arm64-v8a \
  -DANDROID_PLATFORM=android-24 \
  -DANDROID_STL=c++_static \
  -DCMAKE_BUILD_TYPE=Release
cmake --build build-android --target uaii_phone

Ship libuaii_capi.so next to the binary. uaii_phone with no arguments prints the resolved budget. uaii_phone model.q4_k.gguf generates a few tokens from ids 1, 2, 3.

iOS

cmake -S examples/mobile -B build-ios -G Xcode \
  -DCMAKE_SYSTEM_NAME=iOS \
  -DCMAKE_OSX_ARCHITECTURES=arm64 \
  -DCMAKE_OSX_DEPLOYMENT_TARGET=15.0 \
  -DCMAKE_BUILD_TYPE=Release

Embed libuaii_capi.dylib. On device, uaii doctor prints platform: ios or platform: android.