cutensor-cu12

What it is and what it does

cuTENSOR is NVIDIA's high-performance CUDA library for tensor operations on GPUs. It provides optimized implementations of tensor contractions (including just-in-time kernel compilation), reductions, and element-wise operations, with support for mixed-precision compute (FP64 input with FP32 compute, FP32 input with FP16/BF16/TF32 compute) and tensors up to 64 dimensions. It handles arbitrary data layouts and supports operations like tensor permutations, type conversions, and various activation functions.

The package is distributed as platform-specific wheels for CUDA 12 on x86_64, aarch64, and Windows. It has no Python runtime dependencies and is actively maintained. Installation requires CUDA 12 and an NVIDIA GPU; it cannot run on CPU-only systems.

Use it for:

• Accelerate tensor network simulations and quantum computing workloads on NVIDIA GPUs.
• Optimize mixed-precision deep learning operations requiring FP32 inputs with FP16 or BF16 compute.
• Perform high-dimensional tensor contractions in scientific computing and machine learning frameworks.
• Implement custom tensor operations with just-in-time kernel compilation for domain-specific algorithms.
• Execute partial tensor reductions and element-wise operations on large multi-dimensional arrays.

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

cuTENSOR is a CUDA library for high-performance tensor operations on NVIDIA GPUs, supporting contractions, reductions, element-wise operations, and mixed-precision compute on tensors up to 64 dimensions.

Yes, if you have CUDA 12 and an NVIDIA GPU and need high-performance tensor operations. The package is actively maintained, has no security vulnerabilities, and fills a specialized role in GPU-accelerated computing. The proprietary license and CUDA 12 dependency are the main constraints—verify licensing terms for your use case and confirm CUDA 12 availability before installing.

Install

cutensor-cu12 on PyPI

pip install cutensor-cu12

uv add cutensor-cu12

poetry add cutensor-cu12

Installing cutensor-cu12

Before you install

Medium install friction due to platform-specific wheels (x86_64, aarch64, Windows) and CUDA 12 dependency. Package is actively maintained with a recent release.

License in practice

Licensed under NVIDIA Proprietary Software with no SPDX identifier; license treatment is unclear, so review NVIDIA's terms before use in proprietary or redistributed software.

Quickstart

pip install cutensor-cu12

import cutensor

Requires NVIDIA CUDA 12 and an NVIDIA GPU; not usable on CPU-only systems or with other CUDA versions.

Verify before relying

• Whether Python version constraints exist (requires_python is unspecified in metadata)
• Exact scope of NVIDIA Proprietary Software license and redistribution restrictions
• Whether just-in-time kernel compilation requires additional build tools or CUDA toolkit installation

Package facts

Evidence: cutensor_cu12-2.7.0-py3-none-manylinux2014_aarch64.whl; cutensor_cu12-2.7.0-py3-none-manylinux2014_x86_64.whl; cutensor_cu12-2.7.0-py3-none-win_amd64.whl

Keywords: cuda, nvidia, machine learning, tensor network, high-performance computing

Tags

More Scientific/Engineering packages