Fix support for dataloader with None batches

[ethanwharris]

What does this PR do?

Fixes #6753

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?
• [N/A] 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 update the CHANGELOG? (not for typos, docs, test updates, or internal minor changes/refactorings)

PR review

Anyone in the community is free to review the PR once the tests have passed.
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 🙃

[codecov]

Codecov Report

Merging #7342 (108b409) into master (a6aa1a0) will decrease coverage by 5%.
The diff coverage is 100%.

@@           Coverage Diff           @@
##           master   #7342    +/-   ##
=======================================
- Coverage      91%     86%    -5%     
=======================================
  Files         200     200            
  Lines       12916   13210   +294     
=======================================
- Hits        11764   11317   -447     
- Misses       1152    1893   +741     

[tchaton]

LGTM !

[ananthsub]

what happens in distributed training? won't we go out of sync across ranks if the batch is None on some ranks and non-None on others?

[ethanwharris]

I'm not sure. It will be closer to working than it was before, but this use case may still need some work for distributed. The new behaviour (that is, after this fix) should be the same as when returning None from a step

[ananthsub]

There are two options:

• either the trainer runs an all reduce at each step to determine whether all ranks skip. This is a general solution but wastes perf for any users who won't skip batches
• the user runs the all reduce in their lightning module and determines when to skip. This way it's specific to just modules which need to skip batches or steps for any reason.

I prefer the latter approach because I don't want to slow down the trainer for everyone to handle a few corner cases

What do you think @tchaton @SeanNaren @awaelchli @carmocca

[ethanwharris]

Yeah, I think it's fine to make the user responsible for making all ranks skip in sync if they want distributed skipping

[carmocca]

Returning None in any kind of distributed setting or AMP is currently unsupported. We have this (very old) PR to add support but it's currently frozen

#5359

You can have a look at the proposal though