From 6fa1d2e81ee28886725c586edaf384ff4de9e362 Mon Sep 17 00:00:00 2001 From: zhaodewang Date: Thu, 14 Sep 2023 22:51:41 +0800 Subject: [PATCH] add github action. --- .github/workflows/linux-cpu.yml | 28 ++++++ .github/workflows/macos-cpu.yml | 28 ++++++ .github/workflows/windows-cpu.yml | 29 ++++++ README.md | 147 ++++-------------------------- android_build.sh | 7 -- script/android_build.sh | 26 ++++++ script/download_model.sh | 11 +++ script/linux_build.sh | 21 +++++ script/macos_build.sh | 21 +++++ script/windows_build.ps1 | 28 ++++++ 10 files changed, 212 insertions(+), 134 deletions(-) create mode 100644 .github/workflows/linux-cpu.yml create mode 100755 .github/workflows/macos-cpu.yml create mode 100644 .github/workflows/windows-cpu.yml delete mode 100755 android_build.sh create mode 100755 script/android_build.sh create mode 100755 script/download_model.sh create mode 100755 script/linux_build.sh create mode 100755 script/macos_build.sh create mode 100644 script/windows_build.ps1 diff --git a/.github/workflows/linux-cpu.yml b/.github/workflows/linux-cpu.yml new file mode 100644 index 00000000..05d04eab --- /dev/null +++ b/.github/workflows/linux-cpu.yml @@ -0,0 +1,28 @@ +name: linux-cpu +on: + push: + branches: + - master + - 'feature/**' + paths: + - 'src/**' + - '.github/workflows/linux-cpu.yml' + pull_request: + branches: [master] + paths: + - 'src/**' + - '.github/workflows/linux-cpu.yml' + +concurrency: + group: linux-${{ github.ref }} + cancel-in-progress: true +permissions: + contents: read + +jobs: + linux_buil_test: + runs-on: ubuntu-latest + steps: + - uses: actions/checkout@v3 + - name: build + run: ./script/linux_build.sh \ No newline at end of file diff --git a/.github/workflows/macos-cpu.yml b/.github/workflows/macos-cpu.yml new file mode 100755 index 00000000..8ddd9d61 --- /dev/null +++ b/.github/workflows/macos-cpu.yml @@ -0,0 +1,28 @@ +name: macos-cpu +on: + push: + branches: + - master + - 'feature/**' + paths: + - 'src/**' + - '.github/workflows/macos-cpu.yml' + pull_request: + branches: [master] + paths: + - 'src/**' + - '.github/workflows/macos-cpu.yml' + +concurrency: + group: macos-${{ github.ref }} + cancel-in-progress: true +permissions: + contents: read + +jobs: + macos_buil_test: + runs-on: macos-latest + steps: + - uses: actions/checkout@v3 + - name: build + run: ./script/macos_build.sh diff --git a/.github/workflows/windows-cpu.yml b/.github/workflows/windows-cpu.yml new file mode 100644 index 00000000..92e3428d --- /dev/null +++ b/.github/workflows/windows-cpu.yml @@ -0,0 +1,29 @@ +name: windows-cpu +on: + push: + branches: + - master + - 'feature/**' + paths: + - 'src/**' + - '.github/workflows/windows-cpu.yml' + pull_request: + branches: [master] + paths: + - 'src/**' + - '.github/workflows/windows-cpu.yml' + +concurrency: + group: windows-${{ github.ref }} + cancel-in-progress: true +permissions: + contents: read + +jobs: + windows_build_test: + runs-on: windows-latest + steps: + - uses: actions/checkout@v3 + - uses: ilammy/msvc-dev-cmd@v1 + - name: build + run: powershell .\script\windows_build.ps1 \ No newline at end of file diff --git a/README.md b/README.md index db963d1d..738741f2 100644 --- a/README.md +++ b/README.md @@ -14,151 +14,44 @@ | Baichuan2-7B-Chat | [onnx](https://github.com/wangzhaode/llm-export/releases/tag/baichuan2-7b-chat-onnx) | [mnn](https://github.com/wangzhaode/mnn-llm/releases/tag/untagged-6798382d6309a35e20d0) | -## 用法 -### 0. 模型导出与转换 -可以使用[LLMExporter](https://github.com/wangzhaode/LLMExporter)将模型导出为`onnx`格式，然后使用`mnnconvert`转换为`mnn`模型。 +## 构建 -### 1. 下载本项目 -```bash -git clone https://github.com/wangzhaode/mnn-llm.git -``` -### 2. 编译MNN库 -- 克隆MNN项目，最新正式版是2.5.0 -```bash -git clone https://github.com/alibaba/MNN.git -b 2.5.0 -``` - -- 进入MNN项目, 并构建一个Build目录准备编译 -```bash -cd MNN -mkdir build && cd build -``` - -- 正式编译，可选CPU/CUDA/OpenCL三种，推荐有英伟达显卡的选择CUDA，没显卡的选CPU,有AMD显卡的选择OpenCL - -```bash -# CPU only（Suport Linux/Mac/Windows） -cmake -DCMAKE_BUILD_TYPE=Release .. - -# using CUDA(Support Linux) -cmake -DCMAKE_BUILD_TYPE=Release -DMNN_CUDA=ON .. - -# using OPENCL -cmake -DCMAKE_BUILD_TYPE=Release -DMNN_OPENCL=ON -DMNN_USE_SYSTEM_LIB=ON -DMNN_SEP_BUILD=OFF .. - -# start build(support Linux/Mac) -make -j$(nproc) - -# start build(support Windows) -cmake --build . -- /m:8 - -``` - -- 回到mnn-llm - -```bash -cd ../.. -``` - -- 将MNN库的编译结果拷贝到`mnn-llm/libs`目录下 -```bash -# for Linux/Mac -cp -r MNN/include/MNN include -cp MNN/build/libMNN.so libs/ -cp MNN/build/express/*.so libs/ - -# for windows -cp -r MNN/include/MNN include -cp MNN/build/Debug/MNN.dll libs/ -cp MNN/build/Debug/MNN.lib libs/ +### CPU-Only ``` +# linux +./script/linux_build.sh -- 对于Windows，还需要下载一下第三方库pthread，下载[地址](https://gigenet.dl.sourceforge.net/project/pthreads4w/pthreads-w32-2-9-1-release.zip),下载后解压，打开Pre-built.2\lib\x64， 将pthreadVC2.lib文件拷贝到ChatGLM-MNN的libs文件夹。打开Pre-built.2\include,将下面三个.h文件都放到ChatGLM-MNN的include文件夹。对于windows，项目的最终文件结构如下： -```bash -├───libs -│ ├───MNN.dll -│ ├───MNN.lib -│ └───pthreadVC2.lib -├───include -│ ├───cppjieba -│ ├───limonp -│ ├───MNN -│ ├───chat.hpp -│ ├───httplib.h -│ ├───pthread.h -│ ├───sched.h -│ └───semaphore.h -``` - - -### 3. Download Models -从 `github release` 下载模型文件到 `/path/to/ChatGLM-MNN/resource/models`， 如下： -- 对于Linux/Mac -```bash -cd resource/models -# 下载fp16权值模型, 几乎没有精度损失 -./download_models.sh fp16 -# 对于中国用户，可以使用第三方服务加速下载fp16模型 -./download_models.sh fp16 proxy - -# 下载int8权值模型，极少精度损失，推荐使用 -./download_models.sh int8 -# 对于中国用户，可以使用第三方服务加速下载int8模型 -./download_models.sh int8 proxy +# macos +./script/macos_build.sh -# 下载int4权值模型，有一定精度损失 -./download_models.sh int4 -# 对于中国用户，可以使用第三方服务加速下载int4模型 -./download_models.sh int4 proxy -``` -- 对于windows,将上面的`xxx.sh`替换为`xxx.ps1`文件即可，例如： -```powershell -cd resource/models +# windows msvc +./script/windows_build.ps1 -# 下载fp16权值模型, 几乎没有精度损失 -./download_models.ps1 fp16 -# 对于中国用户，可以使用第三方服务加速下载fp16模型 -./download_models.ps1 fp16 proxy +# android +./script/android_build.sh ``` +### CUDA/OPENCL +`TODO` -### 4. Build and Run +### 4. 执行 -##### Mac/Linux/Windows: ```bash -mkdir build && cd build -# for CPU -cmake .. -# for GPU -cmake -D WITH_CUDA=on .. -# for mini memory device -cmake -D BUILD_MINI_MEM_MODE=on .. - -# start build(support Linux/Mac) -make -j$(nproc) - -# start build(support Windows) -cmake --build . -- /m:8 - -# run (for Linux/Mac) +# linux/macos ./cli_demo # cli demo ./web_demo # web ui demo -# run (for Windows) +# windows .\Debug\cli_demo.exe .\Debug\web_demo.exe ``` -##### Android: -``` -mkdir build -cd build -../android_build.sh -make -j8 -``` - ## Reference -- [ChatGLM-6B](https://huggingface.co/THUDM/chatglm-6b) +- [chatglm-6b](https://huggingface.co/THUDM/chatglm-6b) +- [chatglm2-6b](https://huggingface.co/THUDM/chatglm2-6b) +- [codegeex2-6b](https://huggingface.co/THUDM/codegeex2-6b) +- [Baichuan2-7B-Chat](https://huggingface.co/baichuan-inc/Baichuan2-7B-Chat) +- [Qwen-7B-Chat](https://huggingface.co/tangger/Qwen-7B-Chat) - [cpp-httplib](https://github.com/yhirose/cpp-httplib) - [chatgpt-web](https://github.com/xqdoo00o/chatgpt-web) - [cppjieba](https://github.com/yanyiwu/cppjieba) diff --git a/android_build.sh b/android_build.sh deleted file mode 100755 index 2e1cc23d..00000000 --- a/android_build.sh +++ /dev/null @@ -1,7 +0,0 @@ -cmake .. \ --DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake \ --DANDROID_STL=c++_static \ --DANDROID_ABI="arm64-v8a" \ --DANDROID_NATIVE_API_LEVEL=android-21 \ --DCMAKE_BUILD_TYPE=Release \ --DBUILD_FOR_ANDROID=ON diff --git a/script/android_build.sh b/script/android_build.sh new file mode 100755 index 00000000..04a76030 --- /dev/null +++ b/script/android_build.sh @@ -0,0 +1,26 @@ +# 1. clone MNN +git clone https://github.com/alibaba/MNN.git -b 2.6.3 --depth=1 + +# 2. build MNN +cd MNN/project/android +mkdir build +cd build +../build_64.sh +cd ../../../.. + +# 3. copy headers and libs +cp -r MNN/include/MNN include +cp MNN/project/android/build/libMNN.so MNN/project/android/build/libMNN_Express.so libs + +# 4. build mnn-llm android +mkdir android_build +cd android_build +cmake .. \ +-DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake \ +-DANDROID_STL=c++_static \ +-DANDROID_ABI="arm64-v8a" \ +-DANDROID_NATIVE_API_LEVEL=android-21 \ +-DCMAKE_BUILD_TYPE=Release \ +-DBUILD_FOR_ANDROID=ON +make -j4 +cd .. \ No newline at end of file diff --git a/script/download_model.sh b/script/download_model.sh new file mode 100755 index 00000000..ee7bdf15 --- /dev/null +++ b/script/download_model.sh @@ -0,0 +1,11 @@ +model=$1 +mkdir $model +cd $model +# download models +wget -c https://github.com/wangzhaode/mnn-llm/releases/download/$model-mnn/embedding.mnn +wget -c https://github.com/wangzhaode/mnn-llm/releases/download/$model-mnn/lm.mnn +for i in `seq 0 27` +do + wget -c https://github.com/wangzhaode/mnn-llm/releases/download/$model-mnn/block_$i.mnn +done +cd .. \ No newline at end of file diff --git a/script/linux_build.sh b/script/linux_build.sh new file mode 100755 index 00000000..79c87de1 --- /dev/null +++ b/script/linux_build.sh @@ -0,0 +1,21 @@ +# 1. clone MNN +git clone https://github.com/alibaba/MNN.git -b 2.6.3 --depth=1 + +# 2. build MNN +cd MNN +mkdir build +cd build +cmake -DMNN_LOW_MEMORY=ON .. +make -j4 +cd ../.. + +# 3. copy headers and libs +cp -r MNN/include/MNN include +cp MNN/build/libMNN.so MNN/build/express/libMNN_Express.so libs + +# 4. build mnn-llm +mkdir build +cd build +cmake .. +make -j4 +cd .. \ No newline at end of file diff --git a/script/macos_build.sh b/script/macos_build.sh new file mode 100755 index 00000000..c8564913 --- /dev/null +++ b/script/macos_build.sh @@ -0,0 +1,21 @@ +# 1. clone MNN +git clone https://github.com/alibaba/MNN.git -b 2.6.3 --depth=1 + +# 2. build MNN +cd MNN +mkdir build +cd build +cmake -DMNN_LOW_MEMORY=ON .. +make -j4 +cd ../.. + +# 3. copy headers and libs +cp -r MNN/include/MNN include +cp MNN/build/libMNN.dylib MNN/build/express/libMNN_Express.dylib libs + +# 4. build mnn-llm +mkdir build +cd build +cmake .. +make -j4 +cd .. \ No newline at end of file diff --git a/script/windows_build.ps1 b/script/windows_build.ps1 new file mode 100644 index 00000000..e2ab65d2 --- /dev/null +++ b/script/windows_build.ps1 @@ -0,0 +1,28 @@ +# 1. clone MNN +git clone https://github.com/alibaba/MNN.git -b 2.6.3 --depth=1 + +# 2. build MNN +cd MNN +mkdir build +cd build +# cmake -DMNN_LOW_MEMORY=ON .. +cmake .. +cmake --build . -j 4 +cd ../.. + +# 3. copy headers and libs +cp -r MNN/include/MNN include +mv MNN/build/libMNN.lib libs + +# 4. download pthread +wget -Uri https://gigenet.dl.sourceforge.net/project/pthreads4w/pthreads-w32-2-9-1-release.zip -OutFile "pthreads.zip" +Expand-Archive .\pthreads.zip +mv .\pthreads\Pre-built.2\lib\x64\pthreadVC2.lib libs +cp .\pthreads\Pre-built.2\include\*.h .\include\ + +# 5. build mnn-llm +mkdir build +cd build +cmake .. +cmake --build . -j 4 +cd .. \ No newline at end of file