· ReadyStack
Worked example
Real numbers from this tool, line by line.
Six PyTorch channel ties found before conda-forge: that is what the check returned for a 19-line environment.yml of the kind ML engineers and data scientists keep in a vision-training repo, with four errors and two warnings.
| 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 |
The pytorch channel did not disappear, it froze. PyTorch announced in pytorch/pytorch issue 138506 on 2024-10-21 that 2.5 would be the last release published to the pytorch channel on Anaconda, and pointed conda users to conda-forge's pytorch-cpu and pytorch-gpu packages. On 2026-10-12 the anaconda.org API listed pytorch 2.5.1 as the newest build on the pytorch channel and 2.14.1 on conda-forge. torchvision sits at 0.20.1 on the old channel and 0.28.0 on conda-forge; torchaudio at 2.5.1 against 2.11.0.
What a frozen channel costs
A file that lists - pytorch first under channels: keeps solving to torch 2.5.1. CVE-2025-32434, a critical flaw where torch.load with weights_only=True can still run code, affects every torch below 2.6.0 and is fixed only in 2.6.0. The old channel will never carry 2.6.0, so a rebuilt Docker image or a fresh CI runner quietly reinstalls the vulnerable version.
The channel's builds also sat on Anaconda's defaults channel. Anaconda's terms require a paid plan above 200 employees, and the pricing page lists Business at $50 per user per month (read 2026-10-12). conda-forge is free.
Why grep is not enough
grep for "pytorch" in the sample finds four lines and misses two of the six ties. - pytorch under channels: is the frozen channel; the same word under dependencies: is a package name that is fine on conda-forge. pytorch-cuda and cpuonly look harmless but exist only on the old channel, so they pin the solve to it. The nvidia and defaults channels only matter because the pytorch channel needed them. The check reads the YAML block each line sits in and fires the nvidia and defaults rules only when the same file still uses the pytorch channel. On the clean conda-forge version of the same file it returns zero findings.
The twelve rules
The engine has 12 rules: the channel entry, pytorch:: prefixes, -c pytorch on conda, mamba and micromamba command lines in Dockerfiles and CI, a channels: input in setup-miniconda steps, anaconda.org/pytorch URLs in lock files, pytorch-cuda, cpuonly, pytorch-mutex, torch pins below 2.6.0, torchvision or torchaudio pins at old-channel builds, and the nvidia and defaults channels next to pytorch.
Moving the sample file
The finished file lists conda-forge as its only channel with channel_priority: strict, swaps pytorch=2.5.1 and pytorch-cuda=12.4 for pytorch-gpu>=2.6 plus cuda-version=12.6, and drops the pytorch:: prefix from torchvision.
Free and paid
The check is free in VS Code and on the web page: open or paste one file and get every finding with its line and replacement. A $29 key scans every environment, Docker and CI file in the workspace in one pass and writes a dated Markdown migration plan that maps each pytorch-channel line to its conda-forge line.
15 seconds — what it actually does
Questions people ask
What does conda-forge PyTorch Channel Migration Check do?
It reads an environment.yml, Dockerfile, .condarc or CI YAML file and flags every line that still installs PyTorch from the frozen pytorch Anaconda channel: the channel entry, pytorch:: prefixes, -c pytorch flags, pytorch-cuda, cpuonly, anaconda.org/pytorch URLs and torch pins below 2.6.0. Each finding shows its line number and the conda-forge line that replaces it. It has 12 rules.
Who is conda-forge PyTorch Channel Migration Check for?
It is for ML engineers, data scientists and platform teams who install PyTorch with conda, mamba or micromamba and have environment.yml files, Docker images or GitHub Actions jobs written before PyTorch stopped publishing to its Anaconda channel. It suits teams moving to conda-forge or pixi who need to know which repositories still solve to torch 2.5.1.
Why not just grep for pytorch or ask a chatbot?
A grep for pytorch matches every package line, including the correct conda-forge ones, so the real problems drown in noise. A chatbot cannot see your files. This check knows the YAML block a line sits in: '- pytorch' under channels is the frozen channel, under dependencies it is a package. It also knows pytorch-cuda and cpuonly exist only on the old channel.
What is free and what needs a licence key?
Free, with no key: check the open environment.yml, Dockerfile or CI YAML in VS Code, or paste it into the web page, and get every finding with its line and conda-forge replacement. A $29 licence key scans every environment, Docker and CI file in the workspace in one pass and exports a dated Markdown migration plan mapping each pytorch-channel line to conda-forge.
What does staying on the pytorch channel cost?
The pytorch channel stops at torch 2.5.1, and CVE-2025-32434, a critical torch.load remote code execution flaw, is fixed only in 2.6.0, so those environments keep it. The channel's builds also lean on Anaconda's defaults channel, which needs a paid plan above 200 employees; Anaconda Business is $50 per user per month (read 2026-10-12). conda-forge is free.
When does the PyTorch conda situation change next?
No further date is announced. PyTorch's announcement in pytorch/pytorch issue 138506 (2024-10-21) made 2.5 the last release on the pytorch channel. On 2026-10-12 the anaconda.org API listed pytorch 2.5.1 there and pytorch 2.14.1 on conda-forge. New torch releases arrive only on conda-forge, PyPI and download.pytorch.org.
Ask about this tool
One question, answered by the person who built it. Your email only if you want the answer sent.