Remove pytorch 1.10(dev) CI job

[daniellepintz]

What does this PR do?

This job has been failing since it was first introduced in #8133. It is the only job on our CI that consistently fails, and thus is the biggest blocker to having a green CI, which is highly desired. As agreed in the last FB/Grid weekly sync, we will disable this job until there is a release candidate for Pytorch 1.10.

Fixes #9280

Does your PR introduce any breaking changes? If yes, please list them.

Before submitting

• Was this discussed/approved via a GitHub issue? (not for typos and docs)
• Did you read the contributor guideline, Pull Request section?
• Did you make sure your PR does only one thing, instead of bundling different changes together?
• Did you make sure to update the documentation with your changes? (if necessary)
• Did you write any new necessary tests? (not for typos and docs)
• Did you verify new and existing tests pass locally with your changes?
• Did you list all the breaking changes introduced by this pull request?
• Did you update the CHANGELOG? (not for typos, docs, test updates, or internal minor changes/refactorings)

PR review

Anyone in the community is welcome to review the PR.
Before you start reviewing make sure you have read Review guidelines. In short, see the following bullet-list:

• Is this pull request ready for review? (if not, please submit in draft mode)
• Check that all items from Before submitting are resolved
• Make sure the title is self-explanatory and the description concisely explains the PR
• Add labels and milestones (and optionally projects) to the PR so it can be classified

Did you have fun?

Make sure you had fun coding 🙃

[daniellepintz]

I'm confused why this PR is now suddenly running Docker CI jobs like build-XLA and build-CUDA which are failing, but other PRs aren't running these jobs

[Edit] Answer: It only runs on PR's that modify these paths (and on all pushes to master): https://github.com/PyTorchLightning/pytorch-lightning/blob/2b2537d9a064cd437978b111360b904438452145/.github/workflows/ci_dockers.yml#L10-L18
Looks like these Docker jobs will be the next thing to fix.

[Borda]

In other words we don't care about soon future company between PL and PT, meaning actual master nor releasing 1.4.x won't be fully compatible with PT 1.10...
Not happy about this decision but it is your call 😕

[Borda]

Looks like these Docker jobs will be the next thing to fix.

Hope that by fix you mean real fix not dropping it as this PR =)

[daniellepintz]

@Borda

In other words we don't care about soon future company between PL and PT, meaning actual master nor releasing 1.4.x won't be fully compatible with PT 1.10...

The problem is, I would argue that we already don't care about compatibility between PL and PT 1.10, given that this test has literally been failing since it was introduced. The only way to make people care is to make the job required on PRs, and since we don't want to do that since it changes nightly, the only way to keep our CI green is to delete this job. We will add back testing with 1.10 when there is a stable release candidate, at which point we can make it required.

I know it's painful but I don't see a better way, if our goal is a green CI. Do you have any other ideas?

[Borda]

I know it's painful but I don't see a better way, if our goal is a green CI. Do you have any other ideas?

Can we rather just ignore there few failing test with skip if nightly and add todo note so we have more under control

[daniellepintz]

I would be fine with that, if we make the nightly job required. Otherwise, the same thing will just happen again, if people can merge PRs with the nightly job failing.

[Borda]

I would be fine with that, if we make the nightly job required. Otherwise, the same thing will just happen again, if people can merge PRs with the nightly job failing.

Cc: @PyTorchLightning/core-contributors

[ananthsub]

@daniellepintz I share @Borda 's concerns around removing support for testing PyTorch nightly with Lightning. There could be legitimate bugs we ought to surface by having this testing, and for anyone who's running on pytorch main + lightning main, this test coverage is critical.

My understanding is the only 1.10 tests failing are related to quantization with the average observer.

Given the test failures are targeted, and 1.10 RC is imminent, I'd prefer addressing these test failures specifically first.

@Borda historically, have PyTorch + Lightning nightlies have been passing?

[Borda]

@Borda historically, have PyTorch + Lightning nightlies have been passing?

yes for most of the time and if not it was actively resolved on PL side to prevent future incompatibility

[awaelchli]

@daniellepintz
Lightning RC will be released soon and so we propose to mark the job as required now. This would require toggling a setting in Github action by our admin, no PR required. We will enable this soon.

[daniellepintz]

@awaelchli sounds great, closing now!