Installation¶
DaCeML can be installed by using pip install git+https://github.com/spcl/daceml. It is recommended to install the desired version of PyTorch first.
Alternatively, clone the repository and install using:
VENV_PATH='' make install
See Development for more details on the Makefile.
Installing ONNXRuntime¶
DaCeML executes ONNX operators using ONNXRuntime by default. To enable this, a patched version 1 of ONNXRuntime needs to be installed and setup.
ONNXRuntime can be installed from source or from a prebuilt release.
Prebuilt Release (CPU only)¶
The prebuilt release only supports executing ONNX operators with ONNXRuntime on the CPU execution provider.
Note
The CPU-only restriction is not relevant for nodes which have pure SDFG implementations. Nodes with pure implementations can be executed on CPUs, GPUs and FPGAs.
To install, download the prebuilt release and extract it somewhere.
wget https://github.com/orausch/onnxruntime/releases/download/v2/onnxruntime-daceml-patched.tgz
tar -xzf onnxruntime-daceml-patched.tgz
Afterwards, ensure that the environment variable ORT_RELEASE points to the location of the extracted folder.
Build from Source¶
Clone the patched onnxruntime repository somewhere and build it using the following commands.
git checkout add-session-state
./build.sh --build_shared_lib --parallel --config Release
To enable CUDA, add the relevant arguments. For instance:
./build.sh --use_cuda --cuda_version=10.1 --cuda_home=/usr/local/cuda --cudnn_home=/usr/local/cuda --build_shared_lib --parallel --config Release
See onnxruntime/BUILD.md for more details on building ONNXRuntime.
Afterwards, ensure that the environment variable ORT_ROOT points to the root of cloned repository.