Getting Started¶

In this section, we briefly describe how to install cuQuantum which contains both cuStateVec and cuTensorNet.

For C/C++ users, the easiest way to get started is through the installation from conda-forge described below. More detailed explanations on installation, compilation and usage of each component can be found in cuStateVec Guide and cuTensorNet Guide.

For Python users, cuQuantum Python delivers all functionalities of cuStateVec and cuTensorNet with Python APIs. The easiest way to get started with cuQuantum Python is also through installation from conda-forge described below. For more detailed guide on the installation and usage, please refer to cuQuantum Python Guide.

Install cuQuantum from conda-forge¶

If you already have a Conda environment set up (if not, Miniforge/Mambaforge is a great starting point), cuQuantum can be easily installed from the conda-forge channel, via:

conda install -c conda-forge cuquantum

If you need to select the target CUDA version, use the new cuda-version package:

conda install -c conda-forge cuquantum cuda-version=12

The conda solver will install all required dependencies for you. The package is installed to the current $CONDA_PREFIX, so you can simply update the environment variable as follows:

export CUQUANTUM_ROOT=$CONDA_PREFIX

Note

To enable automatic MPI parallelism for cuTensorNet, you can install cuquantum with an MPI from conda-forge, e.g., conda install -c conda-forge cuquantum openmpi. For detailed guide, please refer to cuTensorNet Guide.

Install cuQuantum Python from conda-forge¶

Similar to the cuQuantum installation introduced above, cuQuantum Python can also be easily installed from the conda-forge channel, via:

conda install -c conda-forge cuquantum-python

The conda solver will install all required dependencies for you. Again, the cuda-version metapackage can be used to constrain the CUDA version.

Install cuQuantum from NVIDIA DevZone¶

cuQuantum package can be downloaded or installed from https://developer.nvidia.com/cuQuantum-downloads?target_os=Linux Further, the cuTensorNet component relies upon the functionality of the cuTENSOR package, available for download at https://developer.nvidia.com/cutensor.

For tarball users, once the tarball file is downloaded to CUQUANTUM_ROOT directory, you can unpack it via:

tar zxvf cuquantum-linux-xxxxx-aa.bb.cc.dd-archive.tar.xz

and update the environment variable:

export CUQUANTUM_ROOT=/path/to/where/tarball/is/unpacked

Note

To enable automatic MPI parallelism for cuTensorNet, please refer to cuTensorNet Guide.

For rpm/deb users, We will be using rpm to illustrate some keypoints. For detailed instruction, please refer to the Download page linked above. Please note,

yum -y install cuquantum

The above installation command would install both cuQuantum SDK for CUDA 11.x and CUDA 12.x. CUDA 12.x would be the default. To switch between them, please use update-alternatives

sudo update-alternatives --config cuquantum

There are 2 programs which provide 'cuquantum'.
  Selection    Command
-----------------------------------------------
*+ 1           /usr/src/libcuquantum/12/distributed_interfaces
   2           /usr/src/libcuquantum/11/distributed_interfaces
Enter to keep the current selection[+], or type selection number: 2

If you only need CUDA 11.x or CUDA 12.x, you may perform yum -y install cuquantum-cuda-11 or yum -y install cuquantum-cuda-12.

When using the rpm/deb packages and the system compiler to build a cuQuantum-based program, it is possible to set CUQUANTUM_ROOT and/or CUSTATEVEC_ROOT/CUTENSORNET_ROOT to /usr/. Symlinks to the headers and shared libraries are set up by the package installer.

Note

Fedora 37 was supported starting from CUDA 12.0 Update 1, so only cuquantum-cuda-12 will be supported on Fedora 37.

cuQuantum Dependencies¶

Below we provide a summary of dependencies required for all cuQuantum components:

cuStateVec¶

GPU Architectures	Volta, Turing, Ampere, Ada, Hopper
NVIDIA GPU with Compute Capability	7.0+
CUDA	11.x, 12.x
CPU architectures	x86_64, ppc64le, ARM64
Operating System	Linux
Driver	450.80.02+ (Linux) for CUDA 11.x 525.60.13+ (Linux) for CUDA 12.x

cuTensorNet¶

GPU Architectures	Volta, Turing, Ampere, Ada, Hopper
NVIDIA GPU with Compute Capability	7.0+
CUDA	11.x or 12.x
CPU architectures	x86_64, ppc64le, ARM64
Operating System	Linux
Driver	450.80.02+ (Linux/CUDA 11) or 525.60.13+ (Linux/CUDA 12)
cuTENSOR	1.6.1+

cuQuantum Python¶

	When Building	When Running
Python	3.9+	3.9+
pip	22.3.1+	N/A
setuptools	>=61.0.0	N/A
wheel	>=0.34.0	N/A
Cython	>=0.29.22,<3	N/A
cuStateVec	1.4.0	~=1.3
cuTensorNet	2.2.0	~=2.1
NumPy	N/A	v1.21+
CuPy (see CuPy installation guide)	N/A	v10.0.0+
PyTorch (optional, see PyTorch installation guide)	N/A	v1.10+
Qiskit (optional, see Qiskit installation guide)	N/A	v0.24.0+
Cirq (optional, see Cirq installation guide)	N/A	v0.6.0+
mpi4py (optional, see mpi4py installation guide)	N/A	v3.1.0+

cuQuantum Appliance¶

CUDA	11.x
Driver	470.57.02+ (Linux)