nvidia-cuda-crt

13.2.78 verified Fri May 01 auth: no python

NVIDIA CUDA C Runtime library for Python, distributed as a PyPI package. Version 13.2.78 requires Python >=3. This package provides the CUDA runtime (CRT) as a shared library, enabling GPU-accelerated computations. It is part of the nvidia-pyindex ecosystem and typically installed automatically with CUDA toolkits. Release cadence follows CUDA toolkit releases.

pip install nvidia-cuda-crt

Common errors

error ModuleNotFoundError: No module named 'nvidia' ↓

cause Missing nvidia namespace package; nvidia-cuda-crt requires nvidia package or nvidia-pyindex.

fix

pip install nvidia-pyindex nvidia-cuda-crt

error ImportError: libcuda.so.1: cannot open shared object file ↓

cause CUDA driver not installed or not in LD_LIBRARY_PATH.

fix

Install NVIDIA driver and set LD_LIBRARY_PATH to include /usr/local/cuda/lib64 or /usr/lib/x86_64-linux-gnu/

error AttributeError: module 'nvidia.cuda_crt' has no attribute 'CUDART' ↓

cause Incorrect version of nvidia-cuda-crt; older versions don't export CUDART directly.

fix

pip install --upgrade nvidia-cuda-crt to 13.2.78 or later.

Warnings

breaking Version 13.x dropped support for Python 2.x and older CUDA toolkits. Ensure your CUDA driver and toolkit are compatible with CUDA 13. ↓

fix Upgrade to CUDA Toolkit 13 and use Python >=3.

gotcha This package does not provide the CUDA driver (nvidia-cuda-driver). It only supplies the runtime library. Missing driver will cause runtime errors. ↓

fix Install nvidia-cuda-driver or ensure NVIDIA driver is installed on the system.

gotcha On some systems, the nvidia-cuda-crt package may conflict with system-installed CUDA or other nvidia packages. Use pip install in a virtual environment. ↓

fix Create a virtual environment and install only the required nvidia packages.

Imports

cuda_crt
```
import nvidia.cuda_crt
```
Direct import of the nvidia module. No earlier import pattern.
CUDART
wrong
```
import cuda_crt
```
correct
```
from nvidia.cuda_crt import CUDART
```
Common mistake: forgetting the nvidia namespace prefix.

Quickstart

Verify CUDA runtime is available and check version.

import os
from nvidia.cuda_crt import CUDART

# Check CUDA runtime version
# Requires CUDA environment to be set up
print(CUDART.get_version())