How to create a Conda environment for PyTorch

PyTorch projects often need a Python stack that can change without disturbing notebooks, data tools, or the Conda base environment. A named Conda environment keeps the interpreter, pip, and PyTorch packages together for one training or inference workspace.

Conda creates the isolated prefix, and the activated environment's python -m pip command installs the current PyTorch CPU wheels from the official PyTorch wheel index. That split follows current PyTorch packaging guidance while still letting Conda own activation and dependency isolation.

Use the CPU wheel command for local development, tests, or CPU inference. CUDA, ROCm, and conda-forge packages need a separate platform decision, and old -c pytorch Conda commands no longer match current official PyTorch releases.

Steps to create a PyTorch Conda environment:

Open a terminal where Conda activation works.

If conda activate returns a shell initialization error, initialize Conda for that shell and open a new terminal before continuing.
Related: How to initialize Conda for a shell

Create a named environment with Python and pip.

$ conda create --name torch-lab python=3.12 pip --yes
Channels:
 - conda-forge
Platform: linux-64
Collecting package metadata (repodata.json): done
Solving environment: done

## Package Plan ##

  environment location: /opt/conda/envs/torch-lab

  added / updated specs:
    - pip
    - python=3.12

The following NEW packages will be INSTALLED:

  pip                conda-forge/noarch::pip-26.1.2-pyh8b19718_0
  python             conda-forge/linux-64::python-3.12.13-hd63d673_0_cpython
##### snipped #####
Preparing transaction: done
Verifying transaction: done
Executing transaction: done

Replace torch-lab with the project environment name. Package versions, builds, and channel names depend on the active Conda distribution.

Confirm that the environment exists.

$ conda info --envs

# conda environments:
#
# * -> active
# + -> frozen
base                     /opt/conda
torch-lab                /opt/conda/envs/torch-lab

The new environment should appear with its own path under the Conda installation.
Related: How to list Anaconda environments

Activate the PyTorch environment.

$ conda activate torch-lab
(torch-lab) $

Install PyTorch and the common torchvision and torchaudio companion packages.

(torch-lab) $ python -m pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
Looking in indexes: https://download.pytorch.org/whl/cpu
Collecting torch
  Downloading torch-2.12.1+cpu-cp312-cp312-manylinux_2_28_x86_64.whl (192.3 MB)
Collecting torchvision
  Downloading torchvision-0.27.1+cpu-cp312-cp312-manylinux_2_28_x86_64.whl (1.8 MB)
Collecting torchaudio
  Downloading torchaudio-2.11.0+cpu-cp312-cp312-manylinux_2_28_x86_64.whl (341 kB)
##### snipped #####
Successfully installed filelock-3.29.0 fsspec-2026.4.0 jinja2-3.1.6 numpy-2.4.4 pillow-12.2.0 torch-2.12.1+cpu torchaudio-2.11.0+cpu torchvision-0.27.1+cpu typing-extensions-4.15.0

Use the PyTorch install selector for CUDA or ROCm wheels. Keep python -m pip after activation so the package install targets this Conda environment.

Confirm that torch is installed under the environment path.

(torch-lab) $ python -m pip show torch
Name: torch
Version: 2.12.1+cpu
Summary: Tensors and Dynamic neural networks in Python with strong GPU acceleration
Home-page: https://pytorch.org
License: BSD-3-Clause
Location: /opt/conda/envs/torch-lab/lib/python3.12/site-packages
Requires: filelock, fsspec, jinja2, networkx, setuptools, sympy, typing-extensions
Required-by: torchvision

The Location value should point inside the named Conda environment, not base or the system Python directory.

Run a PyTorch import and tensor smoke test.

(torch-lab) $ python -c 'import torch; print(torch.__version__); print(torch.rand(2, 3))'
2.12.1+cpu
tensor([[0.2052, 0.7758, 0.2776],
        [0.0726, 0.8090, 0.2486]])