cutensor-cu13

What it is and what it does

cuTENSOR is NVIDIA's proprietary CUDA library for tensor primitives—the low-level building block for GPU-accelerated tensor computations. It provides direct tensor contractions (including JIT-compiled kernels), partial and full reductions, and element-wise operations (permutations, type conversions, activations) on tensors up to 64 dimensions. The library supports mixed-precision workflows: FP64 inputs with FP32 compute, FP32 inputs with FP16/BF16/TF32 compute, and complex-times-real operations.

Installation requires a compatible NVIDIA GPU and CUDA 13 runtime; wheels are provided for Linux (x86_64, aarch64) and Windows. The package has no Python runtime dependencies and is actively maintained. It is typically used as a backend for higher-level tensor frameworks or as a direct compute primitive in machine learning and scientific computing pipelines that need fine-grained control over GPU tensor operations.

Use it for:

• Accelerate tensor contraction operations in machine learning models on NVIDIA GPUs with JIT kernel compilation.
• Perform mixed-precision tensor reductions and element-wise operations in scientific computing workflows.
• Implement custom tensor network algorithms requiring direct control over GPU tensor primitives and data layouts.
• Build tensor manipulation kernels for deep learning frameworks that need low-level CUDA tensor support.
• Execute arbitrary tensor permutations and type conversions on GPU with minimal overhead.

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, including contractions, reductions, and element-wise operations on tensors up to 64 dimensions with mixed-precision support.

Yes, if you have an NVIDIA GPU with CUDA 13 and need direct access to high-performance tensor primitives. The library is actively maintained, has no known vulnerabilities, and is the standard CUDA tensor backend for many ML frameworks. However, the proprietary license requires verification of compliance with NVIDIA's terms, and installation friction is moderate due to GPU and CUDA runtime requirements.

Install

cutensor-cu13 on PyPI

pip install cutensor-cu13

uv add cutensor-cu13

poetry add cutensor-cu13

Installing cutensor-cu13

Before you install

Medium install friction: platform-specific wheels for Linux (aarch64, x86_64) and Windows only; requires CUDA 13 and compatible GPU hardware. Package is actively maintained with recent releases.

License in practice

Licensed under NVIDIA Proprietary Software with unclear treatment—not an open-source license. Users should verify compliance with NVIDIA's terms before deployment in production or commercial contexts.

Quickstart

pip install cutensor-cu13

import cutensor
# Use cutensor for tensor operations on NVIDIA GPU

Requires NVIDIA GPU with CUDA 13 support; CUDA toolkit and cuTENSOR runtime libraries must be installed on the system.

Verify before relying

• Exact Python version support (requires_python is unspecified in metadata)
• Whether JIT kernel compilation requires additional build tools or CUDA headers
• Performance characteristics and typical use-case scale (tensor sizes, operation types)

Package facts

Evidence: cutensor_cu13-2.7.0-py3-none-manylinux2014_aarch64.whl; cutensor_cu13-2.7.0-py3-none-manylinux2014_x86_64.whl; cutensor_cu13-2.7.0-py3-none-win_amd64.whl

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

Tags

More Scientific/Engineering packages