LLaMa等LLM语言模型一般使用SentencePiece tokenizer,在端侧部署需要编译和使用其c++版本。
在安卓平台使用NDK编译
CMakeLists.txt需要进行一些修改:
src/CMakeLists.txt如下位置加上log依赖,否则提示android log相关符号不存在。
此外,入口处的CMakeLists.txt加上
set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -s")
可以使得编译后的so大小显著降低,从16.8MB降低到1.9MB。
build.sh参考
#!/bin/bash
SCRIPT_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )"# please set env SDK_ROOT, NDK_ROOT, ANDROID_ABI, MINSDKVERSION# export SDK_ROOT=/root/codes/cmdline-tools/sdk/
# export NDK_ROOT=/root/codes/cmdline-tools/sdk//ndk/24.0.8215888/mkdir -p ${SCRIPT_DIR}/build
cd ${SCRIPT_DIR}/buildif [ -z ${ANDROID_ABI} ] ; thenANDROID_ABI=arm64-v8a
fi
if [ -z ${MINSDKVERSION} ] ; thenMINSDKVERSION=21
fiecho "SDK_ROOT:" $SDK_ROOT
echo "NDK_ROOT:" $NDK_ROOT
echo "ANDROID_ABI:" $ANDROID_ABI
echo "MINSDKVERSION:" $MINSDKVERSION${SDK_ROOT}/cmake/3.18.1/bin/cmake \-DCMAKE_BUILD_TYPE=Release \-DCMAKE_TOOLCHAIN_FILE=${NDK_ROOT}/build/cmake/android.toolchain.cmake \-DANDROID_ABI=${ANDROID_ABI} \-DANDROID_NDK=${NDK_ROOT} \-DANDROID_PLATFORM=android-${MINSDKVERSION} \-DCMAKE_ANDROID_ARCH_ABI=${ANDROID_ABI} \-DCMAKE_ANDROID_NDK=${NDK_ROOT} \-DCMAKE_MAKE_PROGRAM=${SDK_ROOT}/cmake/3.18.1/bin/ninja \-DCMAKE_SYSTEM_NAME=Android \-DCMAKE_SYSTEM_VERSION=${MINSDKVERSION} \-DANDROID_STL=c++_static \-GNinja \..
if [ $? -ne 0 ]; thenecho "ERROR: cmake $TARGET_NAME failed"exit 1
fi${SDK_ROOT}/cmake/3.18.1/bin/ninja
if [ $? -ne 0 ]; thenecho "ERROR: build $TARGET_NAME failed"exit 1
fi
C++API使用方法
https://github.com/google/sentencepiece/blob/master/doc/api.md