Descriptiontl;dr PyTorch will stop publishing Anaconda packages that depend on Anaconda’s default packages due to the high maintenance costs for conda builds which are not justifiable with the ROI we observe today (as seen in download discrepancies between PyPI vs. conda). We are directing users to utilize our official wheel packages from download.pytorch.org or PyPI, or switch to utilizing conda-forge (pytorch) packages if they would like to continue to use conda.
Deprecating PyTorch’s official Anaconda (-c pytorch) channel
Today we are announcing that 2.5 will be the last release of PyTorch that will be published to the pytorch channel on Anaconda. It’s been a long journey of maintenance but we believe that this move will allow the PyTorch team to focus their maintenance efforts on the platform that is most commonly used by our users.
We understand that this move might inconvenience some users but we believe that this change will result in a net positive as we look to double down our efforts to improve the experience for our wheel packages.
As well, we have met with conda-forge maintainers and are looking to address any gaps that may be present in the pytorch-cpu / pytorch-gpu packages on conda-forge in order to make this move as seamless as possible for users. However, please do note that not all gaps, like full Windows support, will be closed and should not be viewed in the short run as a 1:1 offering with our current Anaconda offering.
Maintenance costs vs. User Benefit
Anaconda builds account for more than a half of total maintenance cost in terms of developer time for all of PyTorch packaging despite only accounting for <5% of total downloads. For context, Anaconda has a separate packaging paradigm from pip that requires users to maintain two packaging setups that don’t always align regularly, requiring PyTorch package maintainers to context switch / duplicate package changes between two separate formats.
As well, from a build standpoint conda builds (for pytorch/pytorch) on average take the same time amount (2.9h) as a wheel build, but runs on 6x the hardware at 8x the cost (c5d.4xlarge vs. c5.24xlarge).
To put into perspective it’s important to also understand the discrepancy of Anaconda builds from a download number perspective as well (2.0 release numbers as a baseline):
- Anaconda: 2,367,875 (from condastats)
- PyPI: 61,811,278 (from pypistats)
NOTE: This does not include download.pytorch.org where we host non-standard pip wheel numbers so the discrepancy might be even bigger.
To contrast with the download numbers, here are the percentage of nightly build failures between Wheel (PyPI) vs. Conda for the year 2023-2024 for pytorch, vision and audio combined.
The high maintenance costs of Anaconda builds do not translate into big download numbers, making the ROI argument for continuing to maintain Anaconda builds difficult.
Reporting issues / comments / concerns
We encourage users to create issues on the pytorch/pytorch Github repository if they find issues with our wheel packages.
As well, we are also encouraging users who wish to stay on conda to post issues about conda-forge’s pytorch-cpu / pytorch-gpu packaging on the official conda-forge/pytorch-cpu-feedstock repository.
tl;dr PyTorch will stop publishing Anaconda packages that depend on Anaconda’s default packages due to the high maintenance costs for conda builds which are not justifiable with the ROI we observe today (as seen in download discrepancies between PyPI vs. conda). We are directing users to utilize our official wheel packages from download.pytorch.org or PyPI, or switch to utilizing conda-forge (pytorch) packages if they would like to continue to use conda.
Deprecating PyTorch’s official Anaconda (-c pytorch) channel
Today we are announcing that 2.5 will be the last release of PyTorch that will be published to the pytorch channel on Anaconda. It’s been a long journey of maintenance but we believe that this move will allow the PyTorch team to focus their maintenance efforts on the platform that is most commonly used by our users.
We understand that this move might inconvenience some users but we believe that this change will result in a net positive as we look to double down our efforts to improve the experience for our wheel packages.
As well, we have met with conda-forge maintainers and are looking to address any gaps that may be present in the pytorch-cpu / pytorch-gpu packages on conda-forge in order to make this move as seamless as possible for users. However, please do note that not all gaps, like full Windows support, will be closed and should not be viewed in the short run as a 1:1 offering with our current Anaconda offering.
Maintenance costs vs. User Benefit
Anaconda builds account for more than a half of total maintenance cost in terms of developer time for all of PyTorch packaging despite only accounting for <5% of total downloads. For context, Anaconda has a separate packaging paradigm from pip that requires users to maintain two packaging setups that don’t always align regularly, requiring PyTorch package maintainers to context switch / duplicate package changes between two separate formats.
As well, from a build standpoint conda builds (for pytorch/pytorch) on average take the same time amount (2.9h) as a wheel build, but runs on 6x the hardware at 8x the cost (c5d.4xlarge vs. c5.24xlarge).
To put into perspective it’s important to also understand the discrepancy of Anaconda builds from a download number perspective as well (2.0 release numbers as a baseline):
- Anaconda: 2,367,875 (from condastats)
- PyPI: 61,811,278 (from pypistats)
NOTE: This does not include download.pytorch.org where we host non-standard pip wheel numbers so the discrepancy might be even bigger.
To contrast with the download numbers, here are the percentage of nightly build failures between Wheel (PyPI) vs. Conda for the year 2023-2024 for pytorch, vision and audio combined.
The high maintenance costs of Anaconda builds do not translate into big download numbers, making the ROI argument for continuing to maintain Anaconda builds difficult.
Reporting issues / comments / concerns
We encourage users to create issues on the pytorch/pytorch Github repository if they find issues with our wheel packages.
As well, we are also encouraging users who wish to stay on conda to post issues about conda-forge’s pytorch-cpu / pytorch-gpu packaging on the official conda-forge/pytorch-cpu-feedstock repository.