Build with WASI-NN Plug-in

The WASI-NN plug-in is a proposed WebAssembly System Interface (WASI) API for machine learning. It allows WebAssembly programs to access host-provided machine learning functions.

Prerequisites

Currently, WasmEdge supports following backends for WASI-NN proposal:

Backend	Dependency	CMake Option
OpenVINO	OpenVINO™ (2023)	-DWASMEDGE_PLUGIN_WASI_NN_BACKEND=OpenVINO
TensorFlow-Lite	TensorFlow Lite	-DWASMEDGE_PLUGIN_WASI_NN_BACKEND=TensorFlowLite
PyTorch	PyTorch 1.8.2 LTS	-DWASMEDGE_PLUGIN_WASI_NN_BACKEND=PyTorch
GGML	llama.cpp	-DWASMEDGE_PLUGIN_WASI_NN_BACKEND=GGML
Piper	Piper	-DWASMEDGE_PLUGIN_WASI_NN_BACKEND=Piper
Whisper	whisper.cpp	-DWASMEDGE_PLUGIN_WASI_NN_BACKEND=Whisper
ChatTTS	ChatTTS	-DWASMEDGE_PLUGIN_WASI_NN_BACKEND=ChatTTS
MLX	MLX	-DWASMEDGE_PLUGIN_WASI_NN_BACKEND=MLX

Developers can build the WasmEdge from source with the cmake option WASMEDGE_PLUGIN_WASI_NN_BACKEND to enable the backends. For supporting multiple backends, developers can assign the option such as -DWASMEDGE_PLUGIN_WASI_NN_BACKEND="GGML;Whisper;TensorFlowLite".

After building, you will have the WASI-NN with specified backend(s) plug-in shared library under <YOUR_BUILD_FOLDER>/plugins/wasi_nn/libwasmedgePluginWasiNN.so (or .dylib extension on MacOS).

备注

If the wasmedge CLI tool cannot find the WASI-NN plug-in, you can set the WASMEDGE_PLUGIN_PATH environment variable to the plug-in installation path (such as /usr/local/lib/wasmedge/, or the built plug-in path build/plugins/wasi_nn/) to try to fix this issue.

For the Burn.rs backend, please use the cmake option WASMEDGE_PLUGIN_WASI_NN_BURNRS_MODEL to assign the model.

Model for Burn.rs backend	CMake Option
Squeezenet	-WASMEDGE_PLUGIN_WASI_NN_BURNRS_MODEL=Squeezenet
Whisper	-WASMEDGE_PLUGIN_WASI_NN_BURNRS_MODEL=Whisper

After building, you will have the WASI-NN with specified backend(s) plug-in shared library under <YOUR_BUILD_FOLDER>/plugins/wasi_nn_burnrs/libwasmedgePluginWasiNN.so (or .dylib extension on MacOS).

备注

The WASI-NN Burn.rs backend cannot build with other backends.

Build WasmEdge with WASI-NN OpenVINO Backend

For choosing and installing OpenVINO™ on Ubuntu 20.04 for the backend, we recommend the following commands:

wget https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
sudo apt-key add GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
echo "deb https://apt.repos.intel.com/openvino/2023 ubuntu20 main" | sudo tee /etc/apt/sources.list.d/intel-openvino-2023.list
sudo apt update
sudo apt-get -y install openvino
ldconfig

Then build and install WasmEdge from source:

cd <path/to/your/wasmedge/source/folder>
cmake -GNinja -Bbuild -DCMAKE_BUILD_TYPE=Release -DWASMEDGE_PLUGIN_WASI_NN_BACKEND="OpenVINO"
cmake --build build

Build WasmEdge with WASI-NN PyTorch Backend

For choosing and installing PyTorch on Ubuntu 20.04 for the backend, we recommend the following commands:

export PYTORCH_VERSION="1.8.2"
curl -s -L -O --remote-name-all https://download.pytorch.org/libtorch/lts/1.8/cpu/libtorch-cxx11-abi-shared-with-deps-${PYTORCH_VERSION}%2Bcpu.zip
unzip -q "libtorch-cxx11-abi-shared-with-deps-${PYTORCH_VERSION}%2Bcpu.zip"
rm -f "libtorch-cxx11-abi-shared-with-deps-${PYTORCH_VERSION}%2Bcpu.zip"
export LD_LIBRARY_PATH=$(pwd)/libtorch/lib:${LD_LIBRARY_PATH}
export Torch_DIR=$(pwd)/libtorch

For the legacy operating system such as CentOS 7.6, please use the pre-cxx11-abi version of libtorch instead:

export PYTORCH_VERSION="1.8.2"
curl -s -L -O --remote-name-all https://download.pytorch.org/libtorch/lts/1.8/cpu/libtorch-shared-with-deps-${PYTORCH_VERSION}%2Bcpu.zip
unzip -q "libtorch-shared-with-deps-${PYTORCH_VERSION}%2Bcpu.zip"
rm -f "libtorch-shared-with-deps-${PYTORCH_VERSION}%2Bcpu.zip"
export LD_LIBRARY_PATH=$(pwd)/libtorch/lib:${LD_LIBRARY_PATH}
export Torch_DIR=$(pwd)/libtorch

The PyTorch library will be extracted in the current directory ./libtorch.

Then build and install WasmEdge from source:

cd <path/to/your/wasmedge/source/folder>
cmake -GNinja -Bbuild -DCMAKE_BUILD_TYPE=Release -DWASMEDGE_PLUGIN_WASI_NN_BACKEND="PyTorch"
cmake --build build

Build WasmEdge with WASI-NN TensorFlow-Lite Backend

You can build and install WasmEdge from source directly (on Linux x86_64, Linux aarch64, MacOS x86_64, or MacOS arm64 platforms):

cd <path/to/your/wasmedge/source/folder>
cmake -GNinja -Bbuild -DCMAKE_BUILD_TYPE=Release -DWASMEDGE_PLUGIN_WASI_NN_BACKEND="TensorflowLite"
cmake --build build

Installing the necessary libtensorflowlite_c.so and libtensorflowlite_flex.so on both Ubuntu 20.04 and manylinux2014 for the backend, we recommend the following commands:

curl -s -L -O --remote-name-all https://github.com/second-state/WasmEdge-tensorflow-deps/releases/download/TF-2.12.0-CC/WasmEdge-tensorflow-deps-TFLite-TF-2.12.0-CC-manylinux2014_x86_64.tar.gz
tar -zxf WasmEdge-tensorflow-deps-TFLite-TF-2.12.0-CC-manylinux2014_x86_64.tar.gz
rm -f WasmEdge-tensorflow-deps-TFLite-TF-2.12.0-CC-manylinux2014_x86_64.tar.gz

The shared library will be extracted in the current directory ./libtensorflowlite_c.so and ./libtensorflowlite_flex.so.

Then you can move the library to the installation path:

mv libtensorflowlite_c.so /usr/local/lib
mv libtensorflowlite_flex.so /usr/local/lib

Or set the environment variable export LD_LIBRARY_PATH=$(pwd):${LD_LIBRARY_PATH}.

备注

We also provided the darwin_x86_64, darwin_arm64, and manylinux_aarch64 versions of the TensorFlow-Lite pre-built shared libraries.

For more information, you can refer to the GitHub repository.

Build WasmEdge with WASI-NN llama.cpp Backend

You don't need to install any llama.cpp libraries. WasmEdge will download it during the building period.

Due to the acceleration frameworks being various, you will need to use different compilation options to build this plugin. Please make sure you are following the same OS section to do this.

Build with llama.cpp Backend on MacOS

Intel Model

If you are using the Intel Model macOS, we won't enable any acceleration framework. It is a pure CPU mode plugin.

cd <path/to/your/wasmedge/source/folder>
# Disable BLAS and METAL on x86_64 macOS.
cmake -GNinja -Bbuild -DCMAKE_BUILD_TYPE=Release \
  -DWASMEDGE_PLUGIN_WASI_NN_BACKEND="GGML" \
  -DWASMEDGE_PLUGIN_WASI_NN_GGML_LLAMA_METAL=OFF \
  -DWASMEDGE_PLUGIN_WASI_NN_GGML_LLAMA_BLAS=OFF \
  .
cmake --build build

Apple Silicon Model

You can build and install WasmEdge from source directly on the macOS arm64 platform. It will use the built-in GPU by default.

cd <path/to/your/wasmedge/source/folder>
# Enable METAL on arm64 macOS.
cmake -GNinja -Bbuild -DCMAKE_BUILD_TYPE=Release \
  -DWASMEDGE_PLUGIN_WASI_NN_BACKEND="GGML" \
  -DWASMEDGE_PLUGIN_WASI_NN_GGML_LLAMA_METAL=ON \
  -DWASMEDGE_PLUGIN_WASI_NN_GGML_LLAMA_BLAS=OFF \
  .
cmake --build build

Build with llama.cpp Backend on Linux

Ubuntu/Debian with CUDA 12

Please follow the official guide provided by NVIDIA for installing the CUDA framework: https://developer.nvidia.com/cuda-12-2-0-download-archive

cd <path/to/your/wasmedge/source/folder>

# You may need to install dependencies
apt update
apt install -y software-properties-common lsb-release \
  cmake unzip pkg-config

# Due to cuda-related files, it will produce some warning.
# Disable the warning as an error to avoid failures.
export CXXFLAGS="-Wno-error"
# Please make sure you set up the correct CUDAARCHS.
# We use `60;61;70` for maximum compatibility.
export CUDAARCHS="60;61;70"

# BLAS cannot work with CUBLAS
cmake -GNinja -Bbuild -DCMAKE_BUILD_TYPE=Release \
  -DCMAKE_CUDA_ARCHITECTURES="60;61;70" \
  -DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc \
  -DWASMEDGE_PLUGIN_WASI_NN_BACKEND="GGML" \
  -DWASMEDGE_PLUGIN_WASI_NN_GGML_LLAMA_BLAS=OFF \
  -DWASMEDGE_PLUGIN_WASI_NN_GGML_LLAMA_CUBLAS=ON \
  .

cmake --build build

Ubuntu on NVIDIA Jetson AGX Orin

You should use the pre-built OS image from the NVIDIA official site.

cd <path/to/your/wasmedge/source/folder>

# Due to cuda-related files, it will produce some warning.
# Disable the warning as an error to avoid failures.
export CXXFLAGS="-Wno-error"
# Please make sure you set up the correct CUDAARCHS.
# 72 is for NVIDIA Jetson AGX Orin
export CUDAARCHS=72

# BLAS cannot work with CUBLAS
cmake -GNinja -Bbuild -DCMAKE_BUILD_TYPE=Release \
  -DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc \
  -DWASMEDGE_PLUGIN_WASI_NN_BACKEND="GGML" \
  -DWASMEDGE_PLUGIN_WASI_NN_GGML_LLAMA_BLAS=OFF \
  -DWASMEDGE_PLUGIN_WASI_NN_GGML_LLAMA_CUBLAS=ON \
  .

cmake --build build

Ubuntu/Debian with OpenBLAS

Please install OpenBLAS before building the plugin.

cd <path/to/your/wasmedge/source/folder>

# You may need to install dependencies
apt update
apt install -y software-properties-common lsb-release \
  cmake unzip pkg-config
# You must install OpenBLAS
apt install libopenblas-dev

cmake -GNinja -Bbuild -DCMAKE_BUILD_TYPE=Release \
  -DWASMEDGE_PLUGIN_WASI_NN_BACKEND="GGML" \
  -DWASMEDGE_PLUGIN_WASI_NN_GGML_LLAMA_BLAS=ON \
  .

cmake --build build

General Linux without any acceleration framework

cd <path/to/your/wasmedge/source/folder>

cmake -GNinja -Bbuild -DCMAKE_BUILD_TYPE=Release \
  -DWASMEDGE_PLUGIN_WASI_NN_BACKEND="GGML" \
  -DWASMEDGE_PLUGIN_WASI_NN_GGML_LLAMA_BLAS=OFF \
  .

cmake --build build

Build with llama.cpp Backend on Windows

Install Dependencies for llama.cpp And Build on Windows

Developers can follow the steps for installing the requested dependencies.

1. (Optional, skip this deps if you don't need to use GPU) Download and install CUDA toolkit

   • We use CUDA Toolkit 12 for the release assets
   • Link: https://developer.nvidia.com/cuda-downloads?target_os=Windows&target_arch=x86_64&target_version=11&target_type=exe_local

2. Download and install Visual Studio 2022 Community Edition

   • Link: https://visualstudio.microsoft.com/vs/community/
   • Select the following components in the installer:
     • msvc v143 - vs 2022 c++ x64/x86 build tools (latest)
     • windows 11 sdk (10.0.22621.0)
     • C++ ATL for v143 build tools (x86 & x64)

3. Download and install cmake

   • We use cmake 3.29.3 for the release assets
   • Link: https://github.com/Kitware/CMake/releases/download/v3.29.3/cmake-3.29.3-windows-x86_64.msi

4. Download and install git

   • We use git 2.45.1
   • Link: https://github.com/git-for-windows/git/releases/download/v2.45.1.windows.1/Git-2.45.1-64-bit.exe

5. Download and install ninja-build

   • We use ninja-build 1.12.1
   • Link: https://github.com/ninja-build/ninja/releases/download/v1.12.1/ninja-win.zip
   • Installation: just unzip it to a custom folder

Then developers can build by following the steps.

1. Open Developer PowerShell for VS 2022

   • Start -> Visual Studio 2022 -> Visual Studio Tools -> Developer PowerShell for VS 2022

2. Inside the PowerShell, use git to download wasmedge repo

   cd $HOME
   git clone https://github.com/WasmEdge/WasmEdge.git
   cd WasmEdge

3. Compile wasmedge with enabling the wasi_nn_ggml related options, please use the following commands. To build the plugin, you don't need to enable AOT/LLVM related features, so set them to OFF.

   • If you want to enable CUDA:

     # CUDA ENABLE:
     & "C:\Program files\CMake\bin\cmake.exe" -Bbuild -GNinja -DCMAKE_BUILD_TYPE=Release -DWASMEDGE_PLUGIN_WASI_NN_BACKEND=ggml -DWASMEDGE_PLUGIN_WASI_NN_GGML_LLAMA_CUBLAS=ON -DWASMEDGE_USE_LLVM=OFF .
     & "<the ninja-build folder>\ninja.exe" -C build

   • If you want to disable CUDA:

     # CUDA DISABLE:
     & "C:\Program files\CMake\bin\cmake.exe" -Bbuild -GNinja -DCMAKE_BUILD_TYPE=Release -DWASMEDGE_PLUGIN_WASI_NN_BACKEND=ggml -DWASMEDGE_USE_LLVM=OFF .
     & "<the ninja-build folder>\ninja.exe" -C build

Execute the WASI-NN plugin with the llama example on Windows

1. Set the environment variables

   $env:PATH += ";$pwd\build\lib\api"
   $env:WASMEDGE_PLUGIN_PATH = "$pwd\build\plugins"

2. Download the wasm and run

   wget https://github.com/second-state/WasmEdge-WASINN-examples/raw/master/wasmedge-ggml/llama/wasmedge-ggml-llama.wasm
   wget https://huggingface.co/QuantFactory/Meta-Llama-3-8B-Instruct-GGUF/blob/main/Meta-Llama-3-8B-Instruct.Q5_K_M.gguf
   wasmedge --dir .:. --env llama3=true --env n_gpu_layers=100 --nn-preload default:GGML:AUTO:Meta-Llama-3-8B-Instruct.Q5_K_M.gguf wasmedge-ggml-llama.wasm default

Appendix for llama.cpp backend

We also provided the pre-built ggml plugins on the following platforms:

• darwin_x86_64: Intel Model macOS
• darwin_arm64: Apple Silicon Model macOS
• ubuntu20.04_x86_64: x86_64 Linux (the glibc is using Ubuntu20.04 one)
• ubuntu20.04_aarch64: aarch64 Linux (the glibc is using Ubuntu20.04 one)
• ubuntu20.04_blas_x86_64: x86_64 Linux with OpenBLAS support (the glibc is using Ubuntu20.04 one)
• ubuntu20.04_blas_aarch64: aarch64 Linux with OpenBLAS support (the glibc is using Ubuntu20.04 one)
• ubuntu20.04_cuda_x86_64: x86_64 Linux with CUDA 12 support (the glibc is using Ubuntu20.04 one)
• ubuntu20.04_cuda_aarch64: aarch64 Linux with CUDA 11 support (the glibc is using Ubuntu20.04 one), for NVIDIA Jetson AGX Orin
• manylinux2014_x86_64: x86_64 Linux (the glibc is using CentOS 7 one)
• manylinux2014_aarch64: aarch64 Linux (the glibc is using CentOS 7 one)

Build WasmEdge with WASI-NN Piper Backend

Build and install WasmEdge from source:

cd <path/to/your/wasmedge/source/folder>
cmake -GNinja -Bbuild -DCMAKE_BUILD_TYPE=Release -DWASMEDGE_PLUGIN_WASI_NN_BACKEND="Piper"
cmake --build build

Build WasmEdge with WASI-NN Whisper Backend

Build and install WasmEdge from source:

cd <path/to/your/wasmedge/source/folder>
cmake -GNinja -Bbuild -DCMAKE_BUILD_TYPE=Release -DWASMEDGE_PLUGIN_WASI_NN_BACKEND="Whisper"
cmake --build build

Build WasmEdge with WASI-NN ChatTTS Backend

The ChatTTS backend relies on ChatTTS and Python library, we recommend the following commands to install dependencies.

sudo apt update
sudo apt upgrade
sudo apt install python3-dev
pip install chattts==0.1.1

Then build and install WasmEdge from source:

cd <path/to/your/wasmedge/source/folder>

cmake -GNinja -Bbuild -DCMAKE_BUILD_TYPE=Release -DWASMEDGE_PLUGIN_WASI_NN_BACKEND="chatTTS"
cmake --build build

Build WasmEdge with WASI-NN MLX Backend

You can directly build and install WasmEdge from source or custom install mlx and set CMAKE_INSTALL_PREFIX variable.

Build and install WasmEdge from source:

cd <path/to/your/wasmedge/source/folder>

cmake -GNinja -Bbuild -DCMAKE_BUILD_TYPE=Release -DWASMEDGE_PLUGIN_WASI_NN_BACKEND="mlx"
cmake --build build