Skip to main content

About Pytorch Channel Deprecation

· 3 min read
conda-forge/core
conda-forge/core
The conda-forge core team

PyTorch team announced that they will stop publishing packages to their official channel on Anaconda.org (-c pytorch) after the 2.5 release. We want to address this change and clarify our commitment to maintaining PyTorch packages on conda-forge.

First and foremost, we deeply appreciate the PyTorch team’s significant contributions to the conda packaging ecosystem and their ongoing collaboration with conda-forge. The conda-forge team sees this transition as an opportunity to strengthen the conda ecosystem and create more stable and reliable environments for users.

Dependency conflicts have long been a pain point for users when mixing packages from the conda-forge and Anaconda "defaults." These conflicts arise because different channels build packages against different versions of core dependencies (packages on pytorch channel had dependancies in Anaconda defaults cannels). By consolidating PyTorch packages entirely under conda-forge, we can provide a more consistent and reliable experience, reducing conflicts and improving installation stability.

Our Commitment​

  • Conda-forge is actively working with PyTorch maintainers and the broader open-source community to address potential gaps, ensure stability, and improve package availability.
  • We will continue to maintain pytorch-cpu and pytorch-gpu packages, ensuring they are up to date and aligned with PyTorch releases.
  • While Windows support is currently limited, we are making it a top priority to improve this and deliver a consistent experience across all platforms (see #231).
  • We aim to release updates as close as possible to official PyTorch releases. As a volunteer-driven community, we follow a best-effort approach while prioritizing quality and reliability.
  • We welcome community feedback and contributions, which are essential to improving the ecosystem. Your input directly informs our priorities and helps shape the future of PyTorch on conda-forge.

Current Package Support​

Fully Supported:

  • Linux: Full support for both Pytorch CPU and Pytorch GPU (CUDA versions 11.8, 12.1, and 12.4).
  • macOS: CPU-only support.

Known Limitations:

  • Windows: Limited support currently, but improving Windows support is a top priority.

Migration Guide​

For Current PyTorch Channel Users

Update your conda configuration:

mamba/conda config --remove channels pytorch
mamba/conda config --add channels conda-forge
mamba/conda config --set channel_priority strict

Update your PyTorch installation

mamba install pytorch-cpu
mamba install pytorch-gpu # for GPU support
# OR
conda install pytorch-cpu
conda install pytorch-gpu # for GPU support

Getting Help and Contributing​

The success of this transition depends on collaboration with the community. Here’s how you can get involved:

Frequently Asked Questions​

Will my existing PyTorch installations stop working? No, existing installations will continue to work. However, we recommend transitioning to conda-forge packages for future updates.

How quickly will conda-forge release new PyTorch versions? We aim to release updates as close as possible to official PyTorch releases. But as a volunteer-driven community, we follow a best-effort approach while prioritizing quality and reliability..

What about PyTorch's ecosystem packages (torchvision, torchaudio, etc.)? We maintain these packages as well, and they will continue to be available through conda-forge (see torchvision and torchaudio feedstocks).

How can I ensure package compatibility? Using conda-forge as your primary channel with strict channel priority will ensure package compatibility.