4/n Move Accelerator into strategy - remove X_step() from accelerator

[four4fish]

What does this PR do?

Remove training_step() from Accelerator, and call strategy.training_step directly

In the following graph: Each color is one logic.

[RFC] Should we have training_step() in Parallel Plugin, then DDPSpawning, DP and DDP doesn't need to duplicate the logic. Instead we need training_step in Horovod.
cc: @awaelchli @ananthsub @carmocca @justusschock

Removed TPUSpawning.training_step() as the logic is same as the super class DDPSpawn

Part of #10648

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 🙃

[codecov]

Codecov Report

Merging #10890 (3d6fa8b) into master (6043179) will decrease coverage by 4%.
The diff coverage is 51%.

❗ Current head 3d6fa8b differs from pull request most recent head 23b5e94. Consider uploading reports for the commit 23b5e94 to get more accurate results

@@           Coverage Diff            @@
##           master   #10890    +/-   ##
========================================
- Coverage      92%      88%    -4%     
========================================
  Files         177      177            
  Lines       16484    16570    +86     
========================================
- Hits        15132    14589   -543     
- Misses       1352     1981   +629     

[awaelchli]

[RFC] Should we have training_step() in Parallel Plugin, then DDPSpawning, DP and DDP doesn't need to duplicate the logic. Instead we need training_step in Horovod.

I would keep the overrides for all plugins that wrap the model with a class like DistributedDataParallel for example. This way, the definition of how the model gets wrapped lies in the plugin together with how the training_step gets called through that wrapper.

[four4fish]

[RFC] Should we have training_step() in Parallel Plugin, then DDPSpawning, DP and DDP doesn't need to duplicate the logic. Instead we need training_step in Horovod.

I would keep the overrides for all plugins that wrap the model with a class like DistributedDataParallel for example. This way, the definition of how the model gets wrapped lies in the plugin together with how the training_step gets called through that wrapper.

@awaelchli I agree, even have duplicate logic is better than fail silently or inherent silencely. In the last step of the refactor, we can regrouping logics and flatten the inheritance

[awaelchli]

nice!