Installation

Requirements

Python ≥ 3.14
NumPy ≥ 2.4
pylops ≥ 2.6

These are installed automatically when you install TensorRSVD from PyPI.

From PyPI

pip install tensorrsvd

Optional backends

TensorRSVD ships with optional support for two accelerated backends.

JAX (CPU / GPU / TPU)

JAX enables JIT-compiled matrix–vector products inside MatricizedTensorOperator and is required for the backend="jax" option.

pip install "tensorrsvd[jax]"

For CUDA 12 or CUDA 13 support:

pip install "tensorrsvd[jaxcuda12]"   # CUDA 12
pip install "tensorrsvd[jaxcuda13]"   # CUDA 13

See the JAX installation guide for additional platform-specific instructions.

CuPy (NVIDIA GPU only)

CuPy offloads all linear-algebra primitives to a CUDA device and is required for the backend="cupy" option.

pip install "tensorrsvd[cupy]"

Note

CuPy wheels are CUDA-version-specific. If the above does not match your CUDA installation, install CuPy manually following the CuPy installation guide and choose the wheel that matches your CUDA version.

Verifying the installation

Run a quick smoke test from a Python shell:

import numpy as np
from tensorrsvd import ho_rsvd

def my_tensor(x0, x1, x2):
    return x0 - x1 + x2

U_list, S_list = ho_rsvd(
    tensor=my_tensor,
    tensor_shape=(16, 16, 16),
    dtype=np.float64,
    rank=3,
    num_oversamples=5,
    num_idxs=3,
)
print([U.shape for U in U_list])  # [(16, 3), (16, 3), (16, 3)]

If no exception is raised and the shapes look right, the installation is working correctly.