Lightning-GPU installation

Standard installation

For the majority of cases,

Lightning-GPU can be installed by following our installation instructions at pennylane.ai/install.

Install Lightning-GPU from source

Note

The section below contains instructions for installing Lightning-GPU from source. For most cases, this is not required and one can simply use the installation instructions at pennylane.ai/install. If those instructions do not work for you, or you have a more complex build environment that requires building from source, then consider reading on.

To install Lightning-GPU from source, Lightning-Qubit needs to be ‘installed’ by pip before Lightning-GPU (compilation is not necessary):

git clone https://github.com/PennyLaneAI/pennylane-lightning.git
cd pennylane-lightning
pip install -r requirements.txt
pip install custatevec-cu12
PL_BACKEND="lightning_qubit" python scripts/configure_pyproject_toml.py
SKIP_COMPILATION=True pip install -e . --config-settings editable_mode=compat -vv

Note that custatevec-cu12 is a requirement for Lightning-GPU, and is installed by pip separately. After custatevec-cu12 is installed, the CUQUANTUM_SDK environment variable should be set to enable discovery during installation:

export CUQUANTUM_SDK=$(python -c "import site; print( f'{site.getsitepackages()[0]}/cuquantum')")

The Lightning-GPU can then be installed with pip:

PL_BACKEND="lightning_gpu" python scripts/configure_pyproject_toml.py
python -m pip install -e . --config-settings editable_mode=compat -vv

Lightning-GPU also requires additional NVIDIA libraries including nvJitLink, cuSPARSE, cuBLAS, and CUDA runtime. These can be installed through the CUDA Toolkit or from pip.

To simplify the build, we recommend using the containerized build process described in Docker support section.

Install Lightning-GPU with MPI

Note

Building Lightning-GPU with MPI also requires the NVIDIA cuQuantum SDK (currently supported version: custatevec-cu12), mpi4py and CUDA-aware MPI (Message Passing Interface). CUDA-aware MPI allows data exchange between GPU memory spaces of different nodes without the need for CPU-mediated transfers. Both the MPICH and OpenMPI libraries are supported, provided they are compiled with CUDA support. It is recommended to install the NVIDIA cuQuantum SDK and mpi4py Python package within pip or conda inside a virtual environment. Please consult the cuQuantum SDK , mpi4py, MPICH, or OpenMPI install guide for more information.

Before installing Lightning-GPU with MPI support using the direct SDK path, please ensure that:

Note

• Lightning-Qubit, CUDA-aware MPI and custatevec-cu12 are installed.

• The environment variable CUQUANTUM_SDK is set properly.

• path/to/libmpi.so is added to LD_LIBRARY_PATH.

Then Lightning-GPU with MPI support can be installed in the editable mode:

PL_BACKEND="lightning_gpu" python scripts/configure_pyproject_toml.py
CMAKE_ARGS="-DENABLE_MPI=ON" python -m pip install -e . --config-settings editable_mode=compat -vv

Test Lightning-GPU with MPI

You can test the Python layer of the MPI enabled plugin as follows:

mpirun -np 2 python -m pytest mpitests --tb=short

The C++ code can be tested with:

PL_BACKEND="lightning_gpu" make test-cpp-mpi

lightning_gpu/installation

 

Download Python script

 

Download Notebook

 

View on GitHub