How to Install PyTorch (2026): Windows, Mac, Linux & CUDA

Prerequisites , Before You Install

Before running any install command, confirm your environment meets these requirements. Skipping this step is the single most common reason installation appears to succeed but import torch fails or the wrong Python picks up the package.

pip vs conda: which should you use?

Use pip if you're working in a standard Python virtual environment, a Docker container, or a CI/CD pipeline. It's faster and produces smaller installs. Use conda if you're on Windows with a CUDA GPU , conda bundles the CUDA runtime libraries, so you don't need to install the CUDA toolkit separately. This is a significant time savings for new GPU setups.

Quick Install: CPU-Only (All Platforms)

If you don't have a GPU, or you just want to get started immediately, the CPU-only pip install works on Windows, macOS, and Linux with a single command.

pip install torch torchvision torchaudio

This installs the three core packages: torch (the core framework), torchvision (image datasets and model zoo), and torchaudio (audio processing). All three are versioned together ; don't install them independently or you risk version mismatches.

The CPU-only package is roughly 250 MB. It runs on any hardware including laptops without dedicated GPUs. For learning, prototyping, and inference workloads on small models, CPU-only is sufficient. For training large neural networks, you'll want GPU acceleration , see the CUDA and Apple MPS sections below.

python -c "import torch; x = torch.rand(5, 3); print(x)"

If you see a 5×3 matrix of random numbers, PyTorch is working correctly.

CUDA GPU Installation (Windows & Linux)

GPU-accelerated PyTorch requires matching your CUDA version to the right wheel. PyTorch 2.12.0 supports CUDA 12.6, CUDA 13.0, and CUDA 13.2. The conda approach is recommended for Windows because it bundles the CUDA libraries automatically.

Option 1: conda (recommended for Windows, easy CUDA setup)

conda install pytorch torchvision torchaudio pytorch-cuda=12.6 -c pytorch -c nvidia

The pytorch-cuda=12.6 flag tells conda to install the CUDA 12.6 runtime alongside PyTorch. You still need your NVIDIA GPU driver installed (run nvidia-smi to verify), but you don't need a separate CUDA toolkit installation. This is the fastest path from zero to GPU-accelerated PyTorch on Windows.

Option 2: pip with CUDA wheels

For pip-based CUDA installs, the exact command depends on your CUDA version. Use the official selector at pytorch.org/get-started/locally to generate the correct command. For CUDA 12.6, the general form is:

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126

python -c "import torch; print(torch.cuda.is_available())"
# Expected output: True

python -c "import torch; print(torch.cuda.get_device_name(0))"
# Expected output: your GPU name, e.g. NVIDIA GeForce RTX 4090

If torch.cuda.is_available() returns False after a CUDA install, the most common causes are: NVIDIA driver not installed, CUDA version mismatch between the wheel and your driver, or a virtual environment that picked up a different PyTorch installation. See the Troubleshoot section at the end of this guide.

ROCm for AMD GPUs (Linux only)

PyTorch 2.12.0 supports AMD GPU acceleration via ROCm 7.2 on Linux. Generate the exact ROCm install command from pytorch.org/get-started/locally after selecting your AMD ROCm version. ROCm support is Linux-only ; Windows AMD GPU users should use the CPU build.

macOS Installation (Apple Silicon & Intel)

macOS installation uses pip3 regardless of whether you're on Apple Silicon or Intel. The difference is what GPU acceleration you get afterward.

Apple Silicon (M1/M2/M3/M4) , MPS acceleration

pip3 install torch torchvision torchaudio

On Apple Silicon Macs, this installs a native ARM64 build of PyTorch with Metal Performance Shaders (MPS) support included. MPS lets PyTorch use the unified memory GPU in your M-series chip , no CUDA, no extra configuration. After installing, verify MPS:

python -c "import torch; print(torch.backends.mps.is_available())"
# Expected output: True (on Apple Silicon)

To use MPS in your code, move tensors and models to the MPS device:

device = torch.device("mps" if torch.backends.mps.is_available() else "cpu")
model = MyModel().to(device)
x = x.to(device)

Intel Mac , CPU only

Intel Macs use the same pip3 install torch torchvision torchaudio command. PyTorch does not support CUDA on macOS (NVIDIA stopped CUDA support for macOS in CUDA 10.2). Intel Mac users get CPU-only acceleration. For GPU-accelerated training on macOS hardware, you need an Apple Silicon Mac.

macOS minimum requirement: macOS 10.15 Catalina or higher. Python 3.10–3.14 is required , the system Python on older macOS versions is typically 3.9 or earlier and is not supported.

Linux Installation

Linux is the most fully supported PyTorch platform. All CUDA versions, ROCm, and CPU builds are available. The minimum system requirement is glibc 2.28 or higher, which corresponds to Ubuntu 20.04+, Debian 10+, CentOS 8+, Fedora 24+, and Linux Mint 20+.

CPU-only (any distro)

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu

CUDA on Linux (pip)

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126

CUDA on Linux (conda)

conda install pytorch torchvision torchaudio pytorch-cuda=12.6 -c pytorch -c nvidia

NVIDIA Jetson (edge devices)

For NVIDIA Jetson platforms (Nano, TX1/TX2, Xavier NX/AGX, AGX Orin), PyTorch requires JetPack 4.2 or higher. Install using the NVIDIA-provided wheel for your JetPack version rather than standard pip , the standard wheel will not have the correct ARM library linkages.

Verify Your Installation

Run through these five verification steps in order. Each step builds on the previous , if step 2 fails, there's no point running step 4. Check each one off as you go.

Building from Source (Advanced)

Building from source gives you a PyTorch build customized for your exact hardware and lets you test unreleased features from the main branch. It is significantly more complex than pip or conda and is not recommended for most users.

For source build instructions, refer to the official PyTorch GitHub README. The process involves cloning the repository, installing build dependencies (cmake, ninja, pybind11, etc.), setting environment variables for CUDA paths, and running the build script.

Common Installation Errors Fixed

These are the five errors that account for the majority of PyTorch installation failures. Each has a deterministic fix.