CuPy for CUDA 11.x

13.6.0 verified Mon Apr 27 auth: no python

CuPy is a NumPy/SciPy-compatible array library for GPU-accelerated computing with CUDA 11.x. This package is specifically built for CUDA 11.x environments. Current version: 13.6.0. Release cadence: regular, matching the main CuPy release cycle.

pip install cupy-cuda11x

Common errors

error ImportError: libcudart.so.11.0: cannot open shared object file: No such file or directory ↓

cause The installed cupy-cuda11x package expects CUDA 11.x runtime, but the system has a different or missing CUDA installation.

fix

Install the correct CUDA toolkit version (11.4+) or use a cupy-cuda package matching your CUDA version (e.g., cupy-cuda12x).

error cupy.cuda.runtime.CUDARuntimeError: cudaErrorNoDevice: no CUDA-capable device is detected ↓

cause No NVIDIA GPU is available or the GPU driver is not properly installed.

fix

Check GPU presence with nvidia-smi. Install appropriate NVIDIA drivers.

error ModuleNotFoundError: No module named 'cupy' ↓

cause CuPy has not been installed or the wrong package (e.g., cupy-cuda11x vs cupy) was not installed.

fix

Install with pip install cupy-cuda11x and verify import with import cupy.

Warnings

breaking CuPy v13 drops support for CUDA 11.0-11.3; only CUDA 11.4+ is supported. Ensure your CUDA version is 11.4 or higher. ↓

fix Upgrade CUDA to 11.4+ or use cupy-cuda11x version 12.x for older CUDA 11.x.

gotcha Installing cupy-cuda11x on a system with mismatched CUDA runtime version may cause import errors or silent fallback to CPU. Check with `nvidia-smi` and `nvcc --version`. ↓

fix Ensure that the installed CUDA toolkit version matches the package: CUDA 11.4+ for cupy-cuda11x.

gotcha CuPy arrays created on one GPU device cannot be directly accessed on another device. Use `cp.cuda.Device(n)` context manager to switch devices. ↓

fix Use `with cp.cuda.Device(1):` to perform operations on a different GPU.

deprecated The `cupy.cuda.Device` context management API is deprecated in favor of `cupy.cuda.Device.__enter__` and `__exit__`; use the `with` statement as usual. ↓

fix Use `with cupy.cuda.Device(0):` syntax; it remains the same but is no longer considered experimental.

Imports

cupy
```
import cupy
```
Standard import for CuPy
cupyx
```
import cupyx
```
CuPy extensions (scipy, sparse, etc.)

Quickstart

Basic usage of CuPy on GPU with CUDA 11.x.

import cupy as cp
import numpy as np

# Create a GPU array
x_gpu = cp.array([1, 2, 3, 4, 5])
print(x_gpu)

# Perform operations on GPU
print(cp.sum(x_gpu))

# Create a random matrix on GPU
A = cp.random.randn(1000, 1000)
print(A.shape)

# Verify CUDA and CuPy version
print(cp.__version__)
print(cp.cuda.runtime.runtimeGetVersion())