cuTensorNet (cu12)

24.10.0 verified Fri May 01 auth: no python

cuTensorNet is a component of NVIDIA cuQuantum SDK for tensor network computations, enabling efficient contraction of tensor networks on GPUs. This package (cutensornet-cu12) targets CUDA 12.x. Current version: 24.10.0. It is part of the cuQuantum Python ecosystem, maintained by NVIDIA with quarterly releases.

pip install cutensornet-cu12

Common errors

error ModuleNotFoundError: No module named 'cuquantum' ↓

cause Missing `cuquantum-python-cu12` package (the Python bindings).

fix

Run pip install cuquantum-python-cu12.

error AttributeError: module 'cutensornet' has no attribute 'NetworkOperator' ↓

cause Using wrong import path. The high-level API is in `cuquantum`, not `cutensornet`.

fix

Change to from cuquantum import NetworkOperator.

error TypeError: expected cupy.ndarray, got numpy.ndarray ↓

cause Tensor operands must be GPU arrays (CuPy).

fix

Convert NumPy arrays: cp.asarray(numpy_array).

error cuTensorNet error: CUTENSORNET_STATUS_NOT_SUPPORTED ↓

cause Tensor data type or network topology not supported on the current GPU architecture.

fix

Check cuQuantum compatibility matrix; ensure GPU compute capability >= 7.0 (Volta).

Warnings

breaking The import path for the main API changed from `cutensornet` to `cuquantum` starting in version 22.x. Code using `import cutensornet` will break. ↓

fix Change imports to `from cuquantum import ...`.

gotcha The `cutensornet-cu12` PyPI package only provides the C/CUDA native library. The Python bindings are in `cuquantum-python-cu12`. Installing only `cutensornet-cu12` is insufficient; you must also install `cuquantum-python-cu12`. ↓

fix Install both: `pip install cutensornet-cu12 cuquantum-python-cu12`.

deprecated The `cuquantum` Python package now recommends using `NetworkOperator` instead of the older `einsum` style interface. The old `cutensornet.contract` function is deprecated. ↓

fix Replace `cutensornet.contract(...)` with `NetworkOperator` and `ContractionOptimizer`.

gotcha Tensor operands must be on the GPU (CuPy or similar). Passing NumPy arrays will raise an error. ↓

fix Convert NumPy arrays to CuPy arrays before passing to `NetworkOperator`.

Imports

NetworkOperator
wrong
```
from cutensornet import NetworkOperator
```
correct
```
from cuquantum import NetworkOperator
```
The high-level API is in the `cuquantum` package, not directly in `cutensornet`.
ContractionOptimizer
wrong
```
from cutensornet.cutensornet import ContractionOptimizer
```
correct
```
from cuquantum import ContractionOptimizer
```
The native cuTensorNet C API is exposed via `cuquantum.cutensornet` but users should prefer the high-level API.

Quickstart

Minimal example: define a tensor network, optimize its contraction path, and contract.

import cupy as cp
from cuquantum import NetworkOperator, ContractionOptimizer

# Define tensor network
expr = 'ij,jk,kl->il'
operands = [
    cp.random.rand(2, 3),
    cp.random.rand(3, 4),
    cp.random.rand(4, 2)
]

# Create network operator
net = NetworkOperator(expr, *operands)

# Optimize contraction path
optimizer = ContractionOptimizer(net)
path, cost = optimizer.optimize()

# Contract
tn_result = net.contract(path)
print(tn_result)