conda-forge PyTorch Channel Migration Check
Finds every line that still installs PyTorch from the frozen pytorch Anaconda channel (last torch build there: 2.5.1) and prints the conda-forge line that replaces it. On the sample 19-line environment.yml it finds 6 ties: 4 errors, 2 warnings.
Hub page: https://getreadystack.com/tools/conda-forge-pytorch-channel-check
Why this matters now
On 2024-10-21 PyTorch announced in pytorch/pytorch#138506 that 2.5 would be the last release published to the pytorch channel on Anaconda. The project points conda users to conda-forge (pytorch-cpu / pytorch-gpu) and pip users to download.pytorch.org or PyPI.
Checked on 2026-10-12 with the anaconda.org API:
| package |
pytorch channel (frozen) |
conda-forge |
| pytorch |
2.5.1 |
2.14.1 |
| torchvision |
0.20.1 |
0.28.0 |
| torchaudio |
2.5.1 |
2.11.0 |
| pytorch-cuda |
12.4 |
not on conda-forge, use pytorch-gpu |
| cpuonly |
2.0 |
not on conda-forge, use pytorch-cpu |
An environment that keeps -c pytorch solves to torch 2.5.1 forever. CVE-2025-32434 (critical: torch.load with weights_only=True can run code) affects torch below 2.6.0 and is fixed only in 2.6.0, so the frozen channel can never give you the fix.
What it checks (12 rules)
| rule |
severity |
what it flags |
replacement |
| pytorch-channel-entry |
error |
- pytorch under channels: |
- conda-forge |
| pytorch-channel-prefix |
error |
pytorch::torchvision |
drop the prefix |
| pytorch-channel-cli |
error |
conda install … -c pytorch (also mamba, micromamba) |
-c conda-forge --override-channels |
| setup-miniconda-channels |
error |
channels: conda-forge,pytorch in a workflow |
channels: conda-forge |
| pytorch-channel-url |
error |
conda.anaconda.org/pytorch/ in lock files |
re-lock against conda-forge |
| pytorch-cuda-package |
error |
pytorch-cuda=12.4 |
pytorch-gpu (+ cuda-version) |
| cpuonly-package |
error |
cpuonly |
pytorch-cpu |
| pytorch-mutex-package |
warn |
pytorch-mutex |
delete |
| torch-pin-before-2-6 |
error |
pytorch=2.5.1, torch==2.4.0 |
>=2.6 |
| companion-pin-frozen |
warn |
torchvision=0.20.1, torchaudio==2.5.1 |
unpin or conda-forge builds |
| nvidia-channel-with-pytorch |
warn |
nvidia channel next to pytorch |
delete |
| defaults-with-pytorch |
warn |
defaults channel next to pytorch |
delete, channel_priority: strict |
The last two only fire when the same file also uses the pytorch channel, so a clean conda-forge file stays silent. Comments are skipped.
Worked example (the bundled sample)
| pytorch-channel line |
conda-forge fix |
- pytorch (channels) |
- conda-forge |
- nvidia |
delete: conda-forge carries CUDA |
- defaults |
delete; channel_priority: strict |
- pytorch=2.5.1 |
- pytorch-gpu>=2.6 |
- pytorch::torchvision |
- torchvision |
- pytorch-cuda=12.4 |
- cuda-version=12.6 |
Use
Open an environment.yml, Dockerfile, .condarc or GitHub Actions / GitLab CI YAML file. Findings appear in the Problems panel with the line and the replacement. Run conda-forge PyTorch Check: Check this file from the Command Palette at any time. The same engine runs in the free web page on the hub link above.
Free, no key: the open file, every finding, every replacement line.
Workspace key: scan every environment, Docker and CI file in one pass and export a dated Markdown migration plan.
Yardstick: Anaconda's pricing page lists Business at $50 per user per month (read 2026-10-12); the pytorch channel builds sat on Anaconda's defaults, which needs a paid plan above 200 employees. conda-forge is free.
Sources: pytorch/pytorch#138506 · GHSA-53q9-r3pm-6pq6 · api.anaconda.org (2026-10-12) · anaconda.com/pricing.
