Replace PostLocalSGDOptimizer with a dedicated model averaging component

[daniellepintz]

What does this PR do?

In this PR I am finishing #9446 which was created by @wayi1 before the Strategy refactor. The description below is copied from that PR.

This is an improvement of #8967.

Replace PostLocalSGDOptimizer by a dedicated model averaging component that can run after optimizer_step.

The previous implementation can cause data races and fail the training for some cases. For example, if the data loading phase also involves allreduce (e.g., for checking if the input data stream has reached the end), and such allreduce can be kicked off during optimizer.step(). Therefore, the implementation in this PR can fix such issue.

Proposal: pytorch/pytorch#59699

There are two options to enable post-local SGD by using vanilla PyTorch:

1. Post-LocalSGD Comm Hook + Periodic Model Averager
   https://github.com/pytorch/pytorch/blob/master/torch/distributed/algorithms/model_averaging/averagers.py#L48-L78

   This option requires 1-line addition (such as averager.average_parameters(model.parameters())) in the training loop.

2. Post-LocalSGD Comm Hook + Post-LocalSGD Optimizer
   https://github.com/pytorch/pytorch/blob/master/torch/distributed/optim/post_localSGD_optimizer.py#L15-L48

   This option does not need any code change in the training loop. I tried this option earlier. However, one limitation is that if there is another phase such as data loading that also runs some communication like allreduce and overlaps with optimizer_step(), then this can cause potential data race and fail the training. Since such concurrent phase like data loading is out of the control of model averaging, I didn't choose this implementation.

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 🙃

[awaelchli]

I wonder if this is compatible with ddp spawn. I don't immediately see a reason why not.
Since launchers have been extracted, we want to align the two strategies as much as possible and eventually merge them into one.

[ananthsub]

please add a test and changelog entry as well

[krshrimali]

Thanks, @daniellepintz for the PR. Looks good to me, left a couple of comments for your review. :)

[daniellepintz]

For some reason the tests for PT <1.10 are complaining name 'post_localSGD' is not defined even though the test is marked as min_torch="1.10.0". Seems to be an issue with pytest.

[tchaton]

LGTM !

[ananthsub]

@daniellepintz please also update the docs for DDP optimizations with this support

[daniellepintz]

I updated the docs but something seems wrong with the way it is displaying in the CI. The "note" and code sections get combined: https://153988-178626720-gh.circle-artifacts.com/0/html/advanced/model_parallel.html#ddp-optimizations compared to the current docs: https://pytorch-lightning.readthedocs.io/en/latest/advanced/model_parallel.html?highlight=ddp#ddp-optimizations

however it seems this is an issue in all CI jobs, for example this is the same page from #12245: https://153982-178626720-gh.circle-artifacts.com/0/html/advanced/model_parallel.html#ddp-optimizations

[daniellepintz]

I believe this is ready to be merged - can someone please enable auto merge?

[daniellepintz]

Confirmed the docs on master look good: https://pytorch-lightning.readthedocs.io/en/latest/advanced/model_parallel.html?highlight=ddp#ddp-optimizations

So it was just a weird thing with the CI