安卓端侧大模子 MLC-LLM 摆设全攻略：以 InternLM2.5-1.8B 为例
本文来自社区投稿，作者：TimMLC-LLM 是一个机器学习编译器和高性能大型语言模子摆设引擎。该项目的使命是让每个人都能在自己的平台上开发、优化和摆设 AI 模子。InternLM 2.5 是上海人工智能实验室发布的新一代大规模语言模子，相比于之前的版本，InternLM 2.5支持百万长文，推理能力开源领先。本文将带各人手把手使用 MLC-LLM 将 InternLM2.5-1.8B-Chat摆设到安卓手机上。
https://github.com/InternLM/InternLM
起首我们来看一下最终的结果～
https://i-blog.csdnimg.cn/direct/75dd5404dbaa48738c93be744fec2510.png
1 环境预备
1.1 安装 rust
可参考 Other Installation Methods - Rust Forge
此处使用了国内的镜像，如下列下令，当出现选项时选择 Enter 安装。
export RUSTUP_DIST_SERVER=https://mirrors.ustc.edu.cn/rust-static
export RUSTUP_UPDATE_ROOT=https://mirrors.ustc.edu.cn/rust-static/rustup
curl --proto '=https' --tlsv1.2 -sSf https://mirrors.ustc.edu.cn/misc/rustup-install.sh| sh 1.2 安装 Android Studio
可参考 https://developer.android.com/studio
mkdir -p /root/android && cd /root/android
wget https://redirector.gvt1.com/edgedl/android/studio/ide-zips/2024.1.1.12/android-studio-2024.1.1.12-linux.tar.gz
tar -xvzf android-studio-2024.1.1.12-linux.tar.gz
cd android-studio
wget https://dl.google.com/android/repository/commandlinetools-linux-11076708_latest.zip?hl=zh-cn
unzip commandlinetools-linux-11076708_latest.zip\?hl\=zh-cn
export JAVA_HOME=/root/Downloads/android-studio/jbr
cmdline-tools/bin/sdkmanager "ndk;27.0.12077973" "cmake;3.22.1""platforms;android-34" "build-tools;33.0.1" --sdk_root='sdk' 2 转换模子
2.1 安装 mlc-llm
可参考 https://llm.mlc.ai/docs/install/mlc_llm.html （如果下载很慢可以取消重新运行一下，或者当地下载了之后拷已往）
conda create --name mlc-prebuiltpython=3.11
conda activate mlc-prebuilt
conda install -c conda-forge git-lfs
pip install pytorch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 pytorch-cuda=12.1 transformers sentencepiece protobuf
wget https://github.com/mlc-ai/package/releases/download/v0.9.dev0/mlc_llm_nightly_cu122-0.1.dev1445-cp311-cp311-manylinux_2_28_x86_64.whl
wget https://github.com/mlc-ai/package/releases/download/v0.9.dev0/mlc_ai_nightly_cu122-0.15.dev404-cp311-cp311-manylinux_2_28_x86_64.whl
pip install mlc_ai_nightly_cu122-0.15.dev404-cp311-cp311-manylinux_2_28_x86_64.whl
pip install mlc_llm_nightly_cu122-0.1.dev1445-cp311-cp311-manylinux_2_28_x86_64.whl 测试如下输出说明安装正确
python -c "import mlc_llm; print(mlc_llm)" https://i-blog.csdnimg.cn/direct/902d3696f2374ecba9c3b8cbee4af545.png
克隆项目
git clone https://github.com/mlc-ai/mlc-llm.git
cd mlc-llm
git submodule update --init --recursive 2.2 转换参数
使用 mlc_llm 的 convert_weight 对模子参数进行转换和量化，转换后的参数可以跨平台使用
cd android/MLCChat
export TVM_SOURCE_DIR=/root/android/mlc-llm/3rdparty/tvm
export MLC_LLM_SOURCE_DIR=/root/android/mlc-llm
mlc_llm convert_weight /root/models/internlm2_5-1_8b-chat/ \
--quantization q4f16_1 \
-o dist/internlm2_5-1_8b-chat-q4f16_1-MLC 2.3 生成配置
使用 mlc_llm 的 gen_config 生成 mlc-chat-config.json 并处理 tokenizer
出现提示时输入 y
mlc_llm gen_config /root/models/internlm2_5-1_8b-chat/\
--quantization q4f16_1 --conv-template chatml\
-o dist/internlm2_5-1_8b-chat-q4f16_1-MLC
Do you wish to run the custom code? y 2.4 上传到 HuggingFace
上传这一步需要能访问 HuggingFace，大概需要摆设署理，如果没有署理可以直接在接下来的配置中使用此链接https://huggingface.co/timws/internlm2_5-1_8b-chat-q4f16_1-MLC中的模子（和文档 https://llm.mlc.ai/docs/deploy/android.html#android-sdk 中的转换方法一样）
2.5 (可选) 测试转换的模子
在打包之前可以测试模子结果，需要编译成二进制文件，已成功在个人电脑上运行测试代码。
mlc_llm compile ./dist/internlm2_5-1_8b-chat-q4f16_1-MLC/mlc-chat-config.json \
--device cuda -o dist/libs/internlm2_5-1_8b-chat-q4f16_1-MLC-cuda.so 测试编译的模子是否符合预期，手机端运行的结果和测试结果靠近
from mlc_llm import MLCEngine
# Create engine
engine = MLCEngine(model="./dist/internlm2_5-1_8b-chat-q4f16_1-MLC", model_lib="./dist/libs/internlm2_5-1_8b-chat-q4f16_1-MLC-cuda.so")
# Run chat completion in OpenAI API.
print(engine)
for response in engine.chat.completions.create(
messages=[{"role": "user", "content": "你是谁？"}],
stream=True
):
for choice in response.choices:
print(choice.delta.content, end="", flush=True)
print("\n")
engine.terminate() 3 打包运行
3.1 修改配置文件
修改 mlc-package-config.json，参考如下
{
"device": "android",
"model_list": [
{
"model": "HF://timws/internlm2_5-1_8b-chat-q4f16_1-MLC",
"estimated_vram_bytes": 3980990464,
"model_id": "internlm2_5-1_8b-chat-q4f16_1-MLC"
},
{
"model": "HF://mlc-ai/gemma-2b-it-q4f16_1-MLC",
"model_id": "gemma-2b-q4f16_1-MLC",
"estimated_vram_bytes": 3980990464
}
]
} 3.2 运行打包下令
这一步需要能访问 HuggingFace，大概需要摆设署理
mlc_llm package https://i-blog.csdnimg.cn/direct/923565b044f3458f8fa68bfbb06da8f2.png
3.3 创建签名
cd /root/android/mlc-llm/android/MLCChat
/root/android/android-studio/jbr/bin/keytool -genkey -v -keystore my-release-key.jks -keyalg RSA -keysize 2048 -validity 10000
Enter keystore password:
Re-enter new password:
What is your first and last name?
:Any
What is the name of your organizational unit?
:Any
What is the name of your organization?
:Any
What is the name of your City or Locality?
:Any
What is the name of your State or Province?
:Any
What is the two-letter country code for this unit?
:CN
Is CN=Any, OU=Any, O=Any, L=Any, ST=Any, C=CN correct?
:yes
Generating 2,048 bit RSA key pair and self-signed certificate (SHA256withRSA) with a validity of 10,000 days
for: CN=Any, OU=Any, O=Any, L=Any, ST=Any, C=CN
3.4 修改gradle配置
如果是当地可以 WIFI 或 USB 调试，不消签名；在服务器构建需要签名，修改 app/build.gradle 为如下内容，主要是增长了签名部门，注意确认签名文件的位置。
plugins {
id 'com.android.application'
id 'org.jetbrains.kotlin.android'
}
android {
namespace 'ai.mlc.mlcchat'
compileSdk 34
defaultConfig {
applicationId "ai.mlc.mlcchat"
minSdk 26
targetSdk 33
versionCode 1
versionName "1.0"
testInstrumentationRunner "androidx.test.runner.AndroidJUnitRunner"
vectorDrawables {
useSupportLibrary true
}
}
compileOptions {
sourceCompatibility JavaVersion.VERSION_1_8
targetCompatibility JavaVersion.VERSION_1_8
}
kotlinOptions {
jvmTarget = '1.8'
}
buildFeatures {
compose true
}
composeOptions {
kotlinCompilerExtensionVersion '1.4.3'
}
packagingOptions {
resources {
excludes += '/META-INF/{AL2.0,LGPL2.1}'
}
}
signingConfigs {
release {
storeFile file("/root/android/mlc-llm/android/MLCChat/my-release-key.jks")
storePassword "123456"
keyAlias "mykey"
keyPassword "123456"
}
}
buildTypes {
release {
minifyEnabled false
proguardFiles getDefaultProguardFile('proguard-android-optimize.txt'), 'proguard-rules.pro'
signingConfig signingConfigs.release
}
}
}
dependencies {
implementation project(":mlc4j")
implementation 'androidx.core:core-ktx:1.10.1'
implementation 'androidx.lifecycle:lifecycle-runtime-ktx:2.6.1'
implementation 'androidx.activity:activity-compose:1.7.1'
implementation platform('androidx.compose:compose-bom:2022.10.00')
implementation 'androidx.lifecycle:lifecycle-viewmodel-compose:2.6.1'
implementation 'androidx.compose.ui:ui'
implementation 'androidx.compose.ui:ui-graphics'
implementation 'androidx.compose.ui:ui-tooling-preview'
implementation 'androidx.compose.material3:material3:1.1.0'
implementation 'androidx.compose.material:material-icons-extended'
implementation 'androidx.appcompat:appcompat:1.6.1'
implementation 'androidx.navigation:navigation-compose:2.5.3'
implementation 'com.google.code.gson:gson:2.10.1'
implementation fileTree(dir: 'src/main/libs', include: ['
*.aar', '*
.jar'], exclude: [])
testImplementation 'junit:junit:4.13.2'
androidTestImplementation 'androidx.test.ext:junit:1.1.5'
androidTestImplementation 'androidx.test.espresso:espresso-core:3.5.1'
androidTestImplementation platform('androidx.compose:compose-bom:2022.10.00')
androidTestImplementation 'androidx.compose.ui:ui-test-junit4'
debugImplementation 'androidx.compose.ui:ui-tooling'
debugImplementation 'androidx.compose.ui:ui-test-manifest'
} 3.5 下令行编译
运行编译下令，完成后在 app/build/outputs/apk/release 生成 app-release.apk 安装包，下载得手机上运行，运行 App 需要能访问 HuggingFace 下载模子（参考文档 https://llm.mlc.ai/docs/deploy/android.html#android-sdk 中的 bundle 方法，需要 ADB 刷入模子数据）
./gradlew assembleRelease https://i-blog.csdnimg.cn/direct/033e6c8b8b4e482abe6944f0be293eb5.png
3.6 运行体验
[*] 运行 App 需要能访问 HuggingFce 下载模子
[*] 需要大概 4G 运行内存
[*] 如果运行闪退，很大概是下载不完整可以删除重新下载
https://i-blog.csdnimg.cn/direct/75dd5404dbaa48738c93be744fec2510.png
免责声明：如果侵犯了您的权益，请联系站长，我们会及时删除侵权内容，谢谢合作！更多信息从访问主页：qidao123.com:ToB企服之家，中国第一个企服评测及商务社交产业平台。
页:
[1]