llama.cpp for AOSP

llama.cpp packaged for the Android Open Source Project build system. Used by the AAOSP LLM System Service for local inference.

Structure

platform_external_llamacpp/
├── Android.bp          # Soong build rules
├── jni/
│   └── llm_jni.cpp     # JNI bridge (symlinked from frameworks/base)
├── scripts/
│   ├── sync_upstream.sh    # Pull latest llama.cpp release
│   └── download_model.sh  # Download Qwen 2.5 GGUF models
├── src/                # llama.cpp upstream source (not checked in)
│   ├── include/
│   ├── ggml/
│   ├── src/
│   └── common/
└── README.md

Setup

1. Sync llama.cpp source

cd external/llama.cpp
./scripts/sync_upstream.sh

This downloads the latest llama.cpp release and extracts it to src/.

2. Download a model

# Default: Qwen 2.5 3B (recommended for 8GB+ devices)
./scripts/download_model.sh

# Or specify a tier:
./scripts/download_model.sh --tier high    # Qwen 2.5 7B (12GB+ devices)
./scripts/download_model.sh --tier low     # Qwen 2.5 1.5B (4-8GB devices)
./scripts/download_model.sh --tier minimal # Qwen 2.5 0.5B (<4GB devices)

Models are downloaded to $ANDROID_PRODUCT_OUT/data/local/llm/.

3. Build

# Build the static library + JNI shared library
m libllama libllm_jni

4. Install model on device

adb push $ANDROID_PRODUCT_OUT/data/local/llm/*.gguf /data/local/llm/

Model Selection

The LLM System Service auto-selects the model based on device RAM:

RAM	Model	Size	Context
12GB+	Qwen 2.5 7B Q4_K_M	~4.4 GB	8192
8GB+	Qwen 2.5 3B Q4_K_M	~2.0 GB	4096
4-8GB	Qwen 2.5 1.5B Q4_K_M	~1.1 GB	2048
<4GB	Qwen 2.5 0.5B Q8_0	~0.5 GB	1024

Override via system properties:

adb shell setprop persist.llm.model.path /data/local/llm/custom-model.gguf
adb shell setprop persist.llm.context_size 4096
adb shell setprop persist.llm.gpu_layers 0
adb shell setprop persist.llm.n_threads 4

Updating llama.cpp

# Edit LLAMA_CPP_VERSION in sync_upstream.sh, then:
./scripts/sync_upstream.sh

# Rebuild
m libllama libllm_jni

License

• This packaging: Apache 2.0 (same as AOSP)
• llama.cpp: MIT License
• Qwen 2.5 models: Apache 2.0