This is to avoid build breakages like #26721

@suo do you suggest this is less important than, e,g. namedtensor builds?
I think we need to run this in CI, yes, this is important functionality that we don't want to break

@ilia-cher We weigh master vs. PR-time builds as a balance between how likely all PRs across the project are to break it vs. importance in functionality. Honestly, I may have argued against the inclusion of NamedTensor builds at the time too 😛.

As @ezyang says, it is unlikely that most PRs will break the functionality being tested here, but it will cause potentially extra time waiting + the possibility of flakiness on every PR. I think a decent compromise is what he suggested, where you can manually run the tests by including a special command in your PR description.

@suo TBB build was broken.. I think at least twice in the last few months

In addition, there're external users who specifically rely on these features and these features are not compiled by default.

I can decrease number of builds to alleviate concerns about PR testing time.

Let's find some time to chat after the devcon about this. I appreciate the desire to get this tested, but like I said, I'd like to balance this against the overall developer experience on the project. Typically for other non-default configurations we are content with testing them continuously on master and having the OSS CI oncall revert diffs that break them, rather than testing them against on every PR.

The standard for putting a job in PR time should be: "if this job is failing, does it block the majority of other developers of PyTorch?". In practice, this is not super well-enforced, but we should trying to decrease the number of jobs run on PRs, not increase.

If oncalls are doing their job well reverting diffs then I'm okay :)

cc @zou3519 re namedtensor builds, we can delete them now right?

We can't delete the named tensor builds yet; named tensor tests aren't enabled on the other CI. Once they're enabled on every CI we can delete the named tensor builds.

The plan on the record is to delete the named tensor builds soon.

Updated

@ilia-cher merged this pull request in a22e8f9.

Unfortunately, the newly-added CI builds are failing on master. I am reverting.

Resending #27925

Ok, apparently someone managed to break TBB build while I was landing this PR..

the title of this PR isn't clear at all; the summary is fine, that should be in the title.

the title of this PR isn't clear at all; the summary is fine, that should be in the title.

@gchanan, updated the title of #27925

@ilia-cher could you also update the commit message in that pull request? The commit message is different and most people will be reading the commit title.