Build for On-Device Training
Prerequisites
- Python 3.x
- CMake
Build Instructions for the Training Phase
-
Clone the repository
git clone --recursive https://github.com/Microsoft/onnxruntime.git cd onnxruntime
-
Build ONNX Runtime for
On-Device Training
a. For Windows
.\build.bat --config RelWithDebInfo --cmake_generator "Visual Studio 17 2022" --build_shared_lib --parallel --enable_training_apis
b. For Linux
./build.sh --config RelWithDebInfo --build_shared_lib --parallel --enable_training_apis
c. For Android
Refer to the Android build instructions and add the
--enable_training_apis
build flag.d. For MacOS
Refer to the macOS inference build instructions and add the
--enable_training_apis
build flag.e. For iOS
Refer to the iOS build instructions and add the
--enable_training_apis
build flag.
Note
To build the C# bindings, add the
--build_nuget
flag to the build command above.- To build the Python wheel:
- add the
--build_wheel
flag to the build command above.- install the wheel using
python -m pip install build/Linux/RelWithDebInfo/dist/*.whl
The
config
flag can be one ofDebug
,RelWithDebInfo
,Release
,MinSizeRel
. Use the one that suits your use case.The
--enable_training_apis
flag can be used in conjunction with the--minimal_build
flag.The offline phase of generating the training artifacts can only be done with Python (using the
--build_wheel
flag).- The build commands above only build for the cpu execution provider. To build for cuda execution provider, add these flags
--use_cuda
--cuda_home {directory to your cuda home, for example /usr/local/cuda/}
--cudnn_home {directory to your cuda home, for example /usr/local/cuda/}
--cuda_version={version for example 11.8}
Build for Large Model Training
Contents
Linux
./build.sh --config RelWithDebInfo --build_shared_lib --parallel --enable_training
GPU / CUDA
Prerequisites
The default NVIDIA GPU build requires CUDA runtime libraries installed on the system:
Build instructions
-
Checkout this code repo with
git clone https://github.com/microsoft/onnxruntime cd onnxruntime
- Set the environment variables: adjust the paths for locations on your build machine
export CUDA_HOME=<location for CUDA libs> # e.g. /usr/local/cuda export CUDNN_HOME=<location for cuDNN libs> # e.g. /usr/local/cuda export CUDACXX=<location for NVCC> #e.g. /usr/local/cuda/bin/nvcc
-
Create the ONNX Runtime Python wheel
./build.sh --config=RelWithDebInfo --enable_training --build_wheel --use_cuda --cuda_home {location of cuda libs eg. /usr/local/cuda/} --cudnn_home {location of cudnn libs eg./usr/local/cuda/} --cuda_version={version for eg. 11.8}
-
Install the .whl file in
./build/Linux/RelWithDebInfo/dist
for ONNX Runtime Training.python -m pip install build/Linux/RelWithDebInfo/dist/*.whl
That’s it! Once the build is complete, you should be able to use the ONNX Runtime libraries and executables in your projects. Note that these steps are general and may need to be adjusted based on your specific environment and requirements. For more information, you can ask for help on the ONNX Runtime GitHub community.
GPU / ROCm
Prerequisites
The default AMD GPU build requires ROCm software toolkit installed on the system:
- ROCm 5.4
Build instructions
-
Checkout this code repo with
git clone https://github.com/microsoft/onnxruntime cd onnxruntime
-
Create the ONNX Runtime Python wheel
./build.sh --config Release --enable_training --build_wheel --parallel --skip_tests --use_rocm --rocm_home /opt/rocm
-
Install the .whl file in
./build/Linux/RelWithDebInfo/dist
for ONNX Runtime Training.python -m pip install build/Linux/RelWithDebInfo/dist/*.whl
DNNL and MKLML
Build Instructions
Linux
./build.sh --enable_training --use_dnnl
Windows
.\build.bat --enable_training --use_dnnl
Add --build_wheel
to build the ONNX Runtime wheel.
This will produce a .whl file in build/Linux/RelWithDebInfo/dist
for ONNX Runtime Training.