PyTorch · Module

cuda

torch.cuda provides PyTorch support for CUDA tensor types and GPU computation. Import it safely, then use torch.cuda.is_available() to determine whether CUDA can be used.

How it works

The torch.cuda module provides CUDA-backed tensor support and GPU operations. It can be imported safely because CUDA initialization is lazy; check availability before selecting a CUDA device or allocating CUDA tensors.

Syntax

torch.cuda

Key points

Use the module when a program needs tensors backed by CUDA-capable GPUs.
CUDA tensor operations follow the familiar PyTorch tensor programming model while executing on GPUs.
Moving data and computation to CUDA can enable GPU-based execution instead of CPU execution.
Call torch.cuda.is_available() to test whether the current installation and system can use CUDA.

Common mistakes

Importing the module does not prove that CUDA can be used; call torch.cuda.is_available() before relying on GPU support.

Example: Check availability and create a CUDA tensor

import torch

if torch.cuda.is_available():
    device = torch.device("cuda")
    x = torch.randn(2, 3, device=device)
    print(x.device)
else:
    print("CUDA is unavailable")

Checks CUDA support before creating a tensor on the CUDA device.

When to use it

Runtime capability detection

A program needs to determine whether GPU acceleration is available before choosing a device.

GPU tensor computation

Tensor workloads should execute on a CUDA-capable GPU rather than the CPU.

CUDA resource management

An application needs PyTorch's CUDA-specific device, stream, or initialization utilities.

Interview questions

QWhat is the purpose of torch.cuda?

torch.cuda is PyTorch's CUDA support module. It provides CUDA tensor types and GPU-oriented operations while following the same general tensor-function model as CPU tensors.

Follow-up: How should code handle systems without CUDA? Check torch.cuda.is_available() and select a compatible fallback device before allocating tensors.

QWhen is CUDA initialized, and how can support be checked?

CUDA initialization is lazy, so importing torch.cuda is safe even when the system does not support CUDA. Availability is determined with torch.cuda.is_available().

Follow-up: Why is lazy initialization useful? It lets portable programs import CUDA-related functionality without immediately requiring a usable CUDA runtime.

QHow does a tensor use the GPU through this module?

CUDA tensors use GPUs for computation, so tensors must be allocated on or moved to a CUDA device to run those operations there. A device can be represented with torch.device("cuda").

Follow-up: What should a multi-GPU program add? Select a specific device, such as cuda:0, and manage the active device deliberately.

torch.device
torch.cuda.is_available
torch.cuda.init
torch.cuda.is_initialized
torch.cuda.device
torch.cuda.Stream

Source: torch.cuda — PyTorch 2.8 documentation. Checked 2026-10-16.