cuTensorNet: A High-Performance Library for Tensor Network Computations

Welcome to the cuTensorNet library documentation!

NVIDIA cuTensorNet is a high-performance library for tensor network computations, a component of the NVIDIA cuQuantum SDK. Functionalities of cuTensorNet are described in Overview with installation and usage guide provided in Getting Started.

Key Features

• Based on NVIDIA’s high-performance tensor algebra library: cuTENSOR

• Provides APIs for:

  • Creating a tensor or tensor network

  • Finding a cost-optimal tensor network contraction path for any given tensor network

  • Finding a low-overhead slicing for the tensor network contraction to meet specified memory constraints

  • Tuning the tensor network contraction path finder configuration for better performance

  • Performing tensor network contraction plan generation, auto-tuning, and its subsequent execution

  • Gradually constructing a tensor circuit state (e.g., a quantum circuit state), followed by computing its properties, including arbitrary slices of amplitudes, expectation values, marginal distributions (reduced density matrices), projections on the matrix product state (MPS) space, as well as performing direct sampling of the defined tensor circuit state

  • Compressing the tensor circuit state into the matrix product state (MPS) format

  • Performing backward differentiation (back-propagation) of the tensor network contraction, that is, computing adjoints of the user-specified input tensors, given the adjoint of the output tensor

  • Performing tensor decomposition using QR or SVD

  • Applying a tensor gate (quantum gate) to a pair of connected (contracted) tensors

  • Enabling automatic distributed parallelization in the contraction path finder and executor

  • Enabling custom memory management

  • Logging

Support

• Supported GPU Architectures: Turing, Ampere, Ada, Hopper, Blackwell

• Supported OS: Linux

• Supported CPU Architectures: x86_64, ARM64

Prerequisites

• One of the following CUDA Toolkits and a compatible driver are required:

  CUDA Toolkit	Minimum Required Linux Driver Version
  CUDA® 12.x	>= 525.60.13
  CUDA® 13.x	>= 580.65.06

  Please refer to CUDA Toolkit Release Notes for the details.

• cuTENSOR v2.3.1.

Contents

• Release Notes
  • cuTensorNet v2.9.1
  • cuTensorNet v2.9.0
  • cuTensorNet v2.8.0
  • cuTensorNet v2.7.0
  • cuTensorNet v2.6.0
  • cuTensorNet v2.5.0
  • cuTensorNet v2.4.0
  • cuTensorNet v2.3.0
  • cuTensorNet v2.2.1
  • cuTensorNet v2.2.0
  • cuTensorNet v2.1.0
  • cuTensorNet v2.0.0
  • cuTensorNet v1.1.1
  • cuTensorNet v1.1.0
  • cuTensorNet v1.0.1
  • cuTensorNet v1.0.0
  • cuTensorNet v0.1.0
  • cuTensorNet v0.0.1
• Overview
  • Introduction to tensor networks
  • Migration guide: plan-based → network-centric API (v2.9.0)
  • Contraction optimizer
  • Hyper-optimizer
  • Intermediate tensor(s) reuse
  • Approximation setting
  • Supported data types
  • References
  • Citing cuQuantum
• Examples
  • Compiling code
  • Code example (serial)
  • Code example (automatic slice-based distributed parallelization)
  • Code example (manual slice-based distributed parallelization)
  • Code example (tensorQR)
  • Code example (tensorSVD)
  • Code example (GateSplit)
  • Code example (MPS factorization)
  • Code example (intermediate tensor reuse)
  • Code example (gradients computation)
  • Code example (amplitudes slice)
  • Code example (expectation value)
  • Code example (marginal distribution)
  • Code example (tensor network sampling)
  • Code example (MPS amplitudes slice using simple update)
  • Code example (MPS expectation value)
  • Code example (MPS marginal distribution)
  • Code example (MPS sampling)
  • Code example (MPS sampling QFT)
  • Code example (Projection MPS Circuit DMRG)
  • Code example (MPS sampling MPO)
  • Useful tips
• API Reference
  • cuTensorNet data types
  • cuTensorNet functions
• Acknowledgements