CUDA-Q Simulation Backends¶

The simulation backends that are currently available in CUDA-Q are as follows.

State Vector Simulators¶

The nvidia target provides a state vector simulator accelerated with the cuStateVec library.

The nvidia target supports multiple configurable options.

Features¶

Floating-point precision configuration

The floating point precision of the state vector data can be configured to either double (fp64) or single (fp32) precision. This option can be chosen for the optimal performance and accuracy.

Distributed simulation

The nvidia target supports distributing state vector simulations to multiple GPUs and multiple nodes (mgpu distribution) and multi-QPU (mqpu platform) distribution whereby each QPU is simulated via a single-GPU simulator instance.

Host CPU memory utilization

Host CPU memory can be leveraged in addition to GPU memory to accommodate the state vector (i.e., maximizing the number of qubits to be simulated).

Single-GPU¶

To execute a program on the nvidia target, use the following commands:

Python

python3 program.py [...] --target nvidia

The target can also be defined in the application code by calling

cudaq.set_target('nvidia')

If a target is set in the application code, this target will override the --target command line flag given during program invocation.

C++

nvq++ --target nvidia program.cpp [...] -o program.x
./program.x

By default, this will leverage FP32 floating point types for the simulation. To switch to FP64, specify the --target-option fp64 nvq++ command line option for C++ and Python or use cudaq.set_target('nvidia', option='fp64') for Python in-source target modification instead.

Python

python3 program.py [...] --target nvidia --target-option fp64

The precision of the nvidia target can also be modified in the application code by calling

cudaq.set_target('nvidia', option='fp64')

C++

nvq++ --target nvidia --target-option fp64 program.cpp [...] -o program.x
./program.x

Note

This backend requires an NVIDIA GPU and CUDA runtime libraries. If you do not have these dependencies installed, you may encounter an error stating Invalid simulator requested. See the section Dependencies and Compatibility for more information about how to install dependencies.

In the single-GPU mode, the nvidia target provides the following environment variable options.

**Environment variable options supported in single-GPU mode**¶
Option	Value	Description
`CUDAQ_FUSION_MAX_QUBITS`	positive integer	The max number of qubits used for gate fusion. The default value is `4`.
`CUDAQ_FUSION_DIAGONAL_GATE_MAX_QUBITS`	integer greater than or equal to -1	The max number of qubits used for diagonal gate fusion. The default value is set to `-1` and the fusion size will be automatically adjusted for the better performance. If 0, the gate fusion for diagonal gates is disabled.
`CUDAQ_FUSION_NUM_HOST_THREADS`	positive integer	Number of CPU threads used for circuit processing. The default value is `8`.
`CUDAQ_MAX_CPU_MEMORY_GB`	non-negative integer, or `NONE`	CPU memory size (in GB) allowed for state-vector migration. `NONE` means unlimited (up to physical memory constraints). Default is 0 (disabled).
`CUDAQ_MAX_GPU_MEMORY_GB`	positive integer, or `NONE`	GPU memory (in GB) allowed for on-device state-vector allocation. As the state-vector size exceeds this limit, host memory will be utilized for migration. `NONE` means unlimited (up to physical memory constraints). This is the default.

Deprecated since version 0.8: The nvidia-fp64 targets, which is equivalent setting the fp64 option on the nvidia target, is deprecated and will be removed in a future release.

Multi-node multi-GPU¶

The NVIDIA target also provides a state vector simulator accelerated with the cuStateVec library with support for Multi-Node, Multi-GPU distribution of the state vector, in addition to a single GPU.

The multi-node multi-GPU simulator expects to run within an MPI context. To execute a program on the multi-node multi-GPU NVIDIA target, use the following commands (adjust the value of the -np flag as needed to reflect available GPU resources on your system):

Python

Double precision simulation:

mpiexec -np 2 python3 program.py [...] --target nvidia --target-option fp64,mgpu

Single precision simulation:

mpiexec -np 2 python3 program.py [...] --target nvidia --target-option fp32,mgpu

Note

If you installed CUDA-Q via pip, you will need to install the necessary MPI dependencies separately; please follow the instructions for installing dependencies in the Project Description.

In addition to using MPI in the simulator, you can use it in your application code by installing mpi4py, and invoking the program with the command

mpiexec -np 2 python3 -m mpi4py program.py [...] --target nvidia --target-option fp64,mgpu

The target can also be defined in the application code by calling

cudaq.set_target('nvidia', option='mgpu,fp64')

If a target is set in the application code, this target will override the --target command line flag given during program invocation.

Note

(1) The order of the option settings are interchangeable. For example, cudaq.set_target('nvidia', option='mgpu,fp64') is equivalent to cudaq.set_target('nvidia', option='fp64.mgpu').

The nvidia target has single-precision as the default setting. Thus, using option='mgpu' implies that option='mgpu,fp32'.

C++

Double precision simulation:

nvq++ --target nvidia  --target-option mgpu,fp64 program.cpp [...] -o program.x
mpiexec -np 2 ./program.x

Single precision simulation:

nvq++ --target nvidia  --target-option mgpu,fp32 program.cpp [...] -o program.x
mpiexec -np 2 ./program.x

Note

This backend requires an NVIDIA GPU, CUDA runtime libraries, as well as an MPI installation. If you do not have these dependencies installed, you may encounter either an error stating invalid simulator requested (missing CUDA libraries), or an error along the lines of failed to launch kernel (missing MPI installation). See the section Dependencies and Compatibility for more information about how to install dependencies.

The number of processes and nodes should be always power-of-2.

Host-device state vector migration is also supported in the multi-node multi-GPU configuration.

In addition to those environment variable options supported in the single-GPU mode, the nvidia target provides the following environment variable options particularly for the multi-node multi-GPU configuration.

**Additional environment variable options for multi-node multi-GPU mode**¶
Option	Value	Description
`CUDAQ_MGPU_LIB_MPI`	string	The shared library name for inter-process communication. The default value is `libmpi.so`.
`CUDAQ_MGPU_COMM_PLUGIN_TYPE`	`AUTO`, `EXTERNAL`, `OpenMPI`, or `MPICH`	Selecting `cuStateVec` `CommPlugin` for inter-process communication. The default is `AUTO`. If `EXTERNAL` is selected, `CUDAQ_MGPU_LIB_MPI` should point to an implementation of `cuStateVec` `CommPlugin` interface.
`CUDAQ_MGPU_NQUBITS_THRESH`	positive integer	The qubit count threshold where state vector distribution is activated. Below this threshold, simulation is performed as independent (non-distributed) tasks across all MPI processes for optimal performance. Default is 25.
`CUDAQ_MGPU_FUSE`	positive integer	The max number of qubits used for gate fusion. The default value is `6` if there are more than one MPI processes or `4` otherwise.
`CUDAQ_MGPU_P2P_DEVICE_BITS`	positive integer	Specify the number of GPUs that can communicate by using GPUDirect P2P. Default value is 0 (P2P communication is disabled).
`CUDAQ_GPU_FABRIC`	`MNNVL`, `NVL`, or `NONE`	Automatically set the number of P2P device bits based on the total number of processes when multi-node NVLink (`MNNVL`) is selected; or the number of processes per node when NVLink (`NVL`) is selected; or disable P2P (with `NONE`).
`CUDAQ_GLOBAL_INDEX_BITS`	comma-separated list of positive integers	Specify the inter-node network structure (faster to slower). For example, assuming a 8 nodes, 4 GPUs/node simulation whereby network communication is faster, this `CUDAQ_GLOBAL_INDEX_BITS` environment variable can be set to `3,2`. The first `3` represents 8 nodes with fast communication and the second `2` represents 4 8-node groups in those total 32 nodes. Default is an empty list (no customization based on network structure of the cluster).
`CUDAQ_HOST_DEVICE_MIGRATION_LEVEL`	positive integer	Specify host-device memory migration w.r.t. the network structure.

Deprecated since version 0.8: The nvidia-mgpu target, which is equivalent to the multi-node multi-GPU double-precision option (mgpu,fp64) of the nvidia is deprecated and will be removed in a future release.

The above configuration options of the nvidia backend can be tuned to reduce your simulation runtimes. One of the performance improvements is to fuse multiple gates together during runtime. For example, x(qubit0) and x(qubit1) can be fused together into a single 4x4 matrix operation on the state vector rather than 2 separate 2x2 matrix operations on the state vector. This fusion reduces memory bandwidth on the GPU because the state vector is transferred into and out of memory fewer times. By default, up to 4 gates are fused together for single-GPU simulations, and up to 6 gates are fused together for multi-GPU simulations. The number of gates fused can significantly affect performance of some circuits, so users can override the default fusion level by setting the setting CUDAQ_MGPU_FUSE environment variable to another integer value as shown below.

Python

CUDAQ_MGPU_FUSE=5 mpiexec -np 2 python3 program.py [...] --target nvidia --target-option mgpu,fp64

C++

nvq++ --target nvidia --target-option mgpu,fp64 program.cpp [...] -o program.x
CUDAQ_MGPU_FUSE=5 mpiexec -np 2 ./program.x

OpenMP CPU-only¶

This target provides a state vector simulator based on the CPU-only, OpenMP threaded Q++ library. This is the default target when running on CPU-only systems.

To execute a program on the qpp-cpu target even if a GPU-accelerated backend is available, use the following commands:

Python

python3 program.py [...] --target qpp-cpu

The target can also be defined in the application code by calling

cudaq.set_target('qpp-cpu')

If a target is set in the application code, this target will override the --target command line flag given during program invocation.

C++

nvq++ --target qpp-cpu program.cpp [...] -o program.x
./program.x

Tensor Network Simulators¶

CUDA-Q provides a couple of tensor-network simulator targets accelerated with the cuTensorNet library. These backends are available for use from both C++ and Python.

Tensor network-based simulators are suitable for large-scale simulation of certain classes of quantum circuits involving many qubits beyond the memory limit of state vector based simulators. For example, computing the expectation value of a Hamiltonian via cudaq::observe can be performed efficiently, thanks to cuTensorNet contraction optimization capability. On the other hand, conditional circuits, i.e., those with mid-circuit measurements or reset, despite being supported by both backends, may result in poor performance.

Multi-node multi-GPU¶

The tensornet backend represents quantum states and circuits as tensor networks in an exact form (no approximation). Measurement samples and expectation values are computed via tensor network contractions. This backend supports multi-node, multi-GPU distribution of tensor operations required to evaluate and simulate the circuit.

To execute a program on the tensornet target using a single GPU, use the following commands:

Python

python3 program.py [...] --target tensornet

The target can also be defined in the application code by calling

cudaq.set_target('tensornet')

If a target is set in the application code, this target will override the --target command line flag given during program invocation.

C++

nvq++ --target tensornet program.cpp [...] -o program.x
./program.x

If you have multiple GPUs available on your system, you can use MPI to automatically distribute parallelization across the visible GPUs.

Note

If you installed the CUDA-Q Python wheels, distribution across multiple GPUs is currently not supported for this backend. We will add support for it in future releases. For more information, see this GitHub issue.

Use the following commands to enable distribution across multiple GPUs (adjust the value of the -np flag as needed to reflect available GPU resources on your system):

Python

mpiexec -np 2 python3 program.py [...] --target tensornet

In addition to using MPI in the simulator, you can use it in your application code by installing mpi4py, and invoking the program with the command

mpiexec -np 2 python3 -m mpi4py program.py [...] --target tensornet

C++

nvq++ --target tensornet program.cpp [...] -o program.x
mpiexec -np 2 ./program.x

Note

If the CUTENSORNET_COMM_LIB environment variable is not set, MPI parallelization on the tensornet backend may fail. If you are using a CUDA-Q container, this variable is pre-configured and no additional setup is needed. If you are customizing your installation or have built CUDA-Q from source, please follow the instructions for activating the distributed interface for the cuTensorNet library. This requires installing CUDA development dependencies, and setting the CUTENSORNET_COMM_LIB environment variable to the newly built libcutensornet_distributed_interface_mpi.so library.

Specific aspects of the simulation can be configured by setting the following of environment variables:

`CUDA_VISIBLE_DEVICES=X`: Makes the process only see GPU X on multi-GPU nodes. Each MPI process must only see its own dedicated GPU. For example, if you run 8 MPI processes on a DGX system with 8 GPUs, each MPI process should be assigned its own dedicated GPU via CUDA_VISIBLE_DEVICES when invoking mpiexec (or mpirun) commands.
`OMP_PLACES=cores`: Set this environment variable to improve CPU parallelization.
`OMP_NUM_THREADS=X`: To enable CPU parallelization, set X to NUMBER_OF_CORES_PER_NODE/NUMBER_OF_GPUS_PER_NODE.

Note

This backend requires an NVIDIA GPU and CUDA runtime libraries. If you do not have these dependencies installed, you may encounter an error stating Invalid simulator requested. See the section Dependencies and Compatibility for more information about how to install dependencies.

Note

Setting random seed, via cudaq::set_random_seed, is not supported for this backend due to a limitation of the cuTensorNet library. This will be fixed in future release once this feature becomes available.

Matrix product state¶

The tensornet-mps backend is based on the matrix product state (MPS) representation of the state vector/wave function, exploiting the sparsity in the tensor network via tensor decomposition techniques such as QR and SVD. As such, this backend is an approximate simulator, whereby the number of singular values may be truncated to keep the MPS size tractable. The tensornet-mps backend only supports single-GPU simulation. Its approximate nature allows the tensornet-mps backend to handle a large number of qubits for certain classes of quantum circuits on a relatively small memory footprint.

To execute a program on the tensornet-mps target, use the following commands:

Python

python3 program.py [...] --target tensornet-mps

The target can also be defined in the application code by calling

cudaq.set_target('tensornet-mps')

If a target is set in the application code, this target will override the --target command line flag given during program invocation.

C++

nvq++ --target tensornet-mps program.cpp [...] -o program.x
./program.x

Specific aspects of the simulation can be configured by defining the following environment variables:

`CUDAQ_MPS_MAX_BOND=X`: The maximum number of singular values to keep (fixed extent truncation). Default: 64.
`CUDAQ_MPS_ABS_CUTOFF=X`: The cutoff for the largest singular value during truncation. Eigenvalues that are smaller will be trimmed out. Default: 1e-5.
`CUDAQ_MPS_RELATIVE_CUTOFF=X`: The cutoff for the maximal singular value relative to the largest eigenvalue. Eigenvalues that are smaller than this fraction of the largest singular value will be trimmed out. Default: 1e-5
`CUDAQ_MPS_SVD_ALGO=X`: The SVD algorithm to use. Valid values are: GESVD (QR algorithm), GESVDJ (Jacobi method), GESVDP (polar decomposition), GESVDR (randomized methods). Default: GESVDJ.

Note

This backend requires an NVIDIA GPU and CUDA runtime libraries. If you do not have these dependencies installed, you may encounter an error stating Invalid simulator requested. See the section Dependencies and Compatibility for more information about how to install dependencies.

Note

Setting random seed, via cudaq::set_random_seed, is not supported for this backend due to a limitation of the cuTensorNet library. This will be fixed in future release once this feature becomes available.

Note

The parallelism of Jacobi method (the default CUDAQ_MPS_SVD_ALGO setting) gives GPU better performance on small and medium size matrices. If you expect the a large number of singular values (e.g., increasing the CUDAQ_MPS_MAX_BOND setting), please adjust the CUDAQ_MPS_SVD_ALGO setting accordingly.

Default Simulator¶

If no explicit target is set, i.e. if the code is compiled without any --target flags, then CUDA-Q makes a default choice for the simulator.

If an NVIDIA GPU and CUDA runtime libraries are available, the default target is set to nvidia. This will utilize the cuQuantum single-GPU state vector simulator. On CPU-only systems, the default target is set to qpp-cpu which uses the OpenMP CPU-only simulator.

The default simulator can be overridden by the environment variable CUDAQ_DEFAULT_SIMULATOR. If no target is explicitly specified and the environment variable has a valid value, then it will take effect. This environment variable can be set to any non-hardware backend. Any invalid value is ignored.

For CUDA-Q Python API, the environment variable at the time when cudaq module is imported is relevant, not the value of the environment variable at the time when the simulator is invoked.

For example,

Python

CUDAQ_DEFAULT_SIMULATOR=density-matrix-cpu python3 program.py [...]

C++

CUDAQ_DEFAULT_SIMULATOR=density-matrix-cpu nvq++ program.cpp [...] -o program.x
./program.x

This will use the density matrix simulator target.

Note

To use targets that require an NVIDIA GPU and CUDA runtime libraries, the dependencies must be installed, else you may encounter an error stating Invalid simulator requested. See the section Dependencies and Compatibility for more information about how to install dependencies.