torch.cuda.is_available() returns False

The most common cause is a CPU-only PyTorch build, which is what Windows gets by default. The others are the driver, the environment, or on AMD, permissions. One command tells you which.

Diagnose it first

python -c "import torch; print(torch.__version__, torch.version.cuda, torch.version.hip, torch.cuda.is_available())"

What it prints	What it means
`2.x.x+cpu None None False`, or `None None` with no suffix	A CPU-only build. Go to cause 1.
a CUDA version (e.g. `12.6`) and `False`	A GPU build that can't reach the GPU: driver, environment or container.
a HIP version (e.g. `7.2...`) and `False`	An AMD ROCm build that can't see the GPU. Go to AMD.

1. You installed a CPU-only build

Calling CUDA functions on it raises AssertionError: Torch not compiled with CUDA enabled.On Windows, a plain pip install torch from PyPI always gives you this: PyPI's Windows wheels are CPU-only (their CUDA dependencies are declared for Linux only), so the GPU build has to come from PyTorch's own index:

pip uninstall torch
pip install torch==2.14.0 --index-url https://download.pytorch.org/whl/cu126

Pick the CUDA build that matches your GPU and driver with the install checker: an older GPU may need a specific build, and a newer one may need a newer build than cu126. On Linux, a plain pip install torch does get a CUDA build, but some tools and environments pull the +cpu one; uninstall and reinstall from the right index. macOS has no CUDA at all (Apple GPUs use the mps backend).

2. The NVIDIA driver is missing or too old

Run nvidia-smi. If the command fails, no driver is loaded; install one from NVIDIA (PyTorch reports this as Found no NVIDIA driver on your system). If it works, compare the driver with what your build needs: PyTorch warns that The NVIDIA driver on your system is too old andis_available() returns False. See the driver error page for the numbers.

3. The GPU is hidden from the process

CUDA_VISIBLE_DEVICES set to an empty string or to device numbers that don't exist hides every GPU. Check with echo $CUDA_VISIBLE_DEVICES.
Under WSL2, install the NVIDIA driver on Windows only. Installing a Linux NVIDIA driver inside WSL2 breaks GPU access.

4. You're inside a container without GPU access

Docker containers only see the GPU when started with --gpus all (or the Compose / Kubernetes equivalent) and the NVIDIA Container Toolkit is installed on the host. The CUDA Toolkit itself only needs to be inside the image; the host needs just the driver and the toolkit.

5. AMD: a ROCm build that can't see the GPU

ROCm builds use the same torch.cuda API, so is_available() is the right check.
Your user needs access to the GPU device files, which ROCm's install guide grants through the render and video groups. Check with groups.
Run rocminfo: if your GPU's gfx name isn't listed, the driver or ROCm installation is the problem, not PyTorch.
If the GPU is visible but its gfx target isn't one ROCm or the PyTorch build supports, see hipErrorNoBinaryForGpu and ROCm on consumer GPUs.

Sources: PyPI's Windows wheel metadata for torch 2.14.0; error text from PyTorch's torch/cuda/__init__.py and c10/cuda/CUDAFunctions.cpp; WSL2 and container requirements from NVIDIA's CUDA on WSL and Container Toolkit guides (see CUDA versions).

← All error fixes