PyTorch LTS page: what's the story for torchvision, text, audio, etc.?

[chrish42]

The only page talking about PyTorch LTS / Enterprise Support Program does not talk at all about what happens to torchvision, etc. with the LTS program. Currently, versions for torchvision, etc. have an exact requirements for a corresponding PyTorch version. For example, torch 1.8.0 is tied to torchvision 0.9.0. What's the plan there?

To be specific, the first PyTorch release of the LTS program (1.8.2) has been published. But there's no corresponding release of torchvision, so folks using the latter can't upgrade to 1.8.2 currently. Is that just an oversight and there'll be corresponding releases of torchvision, torchtext and torchaudio for every LTS release? Will the version requirements be relaxed so torchvision 0.9 can be used with any bugfix version of PyTorch LTS 1.8.x (since they're supposed not to have any breaking changes, as LTS bugfix releases)? Or will those be made available only for customers who pay for LTS support? Or anything else?

That webpage could really benefit from a clarification there. At least, for people using torchvision and friends, a LTS track for PyTorch will be of limited use if they can't use torchvision, etc. with it.

The page on the website: https://pytorch.org/enterprise-support-program

[malfet]

cc: @harishsk

[harishsk]

Please see this page for Microsoft's support boundaries for Pytorch LTS: https://docs.microsoft.com/en-us/azure/pytorch-enterprise/support-boundaries

Torchvision, etc. are supported. When 1.8.2 was released, torchvision was updated to 0.9.2. (Please see https://anaconda.org/pytorch-lts )

[gkowarzyk]

Hello @harishsk, following up on this topic.

As you explain, and based on Microsoft's support document, my understanding is that:

• torchvision
• torchaudio
• torchtext are all supported in the context of PyTorch LTS.

More specifically, it states that we support the specific versions that are compatible with the corresponding supported PyTorch version, and provides links to the tables that show a compatibility matrix.

The current LTS version of PyTorch is PyTorch v1.8.2.

The issue we currently have is that none of the tables currently list a specific compatible version matching v1.8.2, see:

• https://github.com/pytorch/vision#installation
• https://github.com/pytorch/audio/#dependencies
• https://github.com/pytorch/text#installation

For torchvision only the version pinned to PyTorch 1.8.0 is listed. For torchtext and torchaudio, only the general matching minor version is listed (but not the exact patch version).

Nevertheless, we see updated tags (but no corresponding official release) in each repo:

• for torchvision, 0.9.2-rc2 is present but no official 0.9.2
• for torchaudio, 0.8.2-rc3 is present but no official 0.8.2
• for torchtext, 0.9.2-rc1 is present but no official 0.9.2
• none of them have an official release in the Releases page. As a consequence, we do not know what the commit should be for the official releases.

We are building our packages from source, and with official releases we can use the given tag/commit to build the recommended version. We understand that there are torchvision / torchaudio / torchtext binary packages for PyTorch v1.8.2 in the anaconda distribution, but we would like to be able to easily know what specific tag / commit they correspond to (what version of the source code was used to build them).

In the context of LTS releases, would it be possible for the different teams to coordinate and:

• keep the compatibility table matrix up to date when new LTS releases are created.
• create a an official Release entry in the Releases page
• ensure that there is a matching tag in the github repo?

Thanks in advance!