Fix wrong behavior of DDPStrategy option with simple GAN training using DDP

[samsara-ku]

What does this PR do?

Fixes #20866
Fixes #20328
Fixes #18740
Fixes #17212

This PR adds MultiModelDDPStrategy class and its simple execution example, for the multi-gpu training with GAN training.

Simply speaking:

* right way using DDP in pytorch.
# each model is divided and wrapped by DDP class
generator, discriminator = DDP(generator), DDP(discrimiantor)

* current pytorch-lightning GAN training.
# trainer contains two nn.Modules; generator and discriminator
DDP(trainer) -> trainer contains two nn.Modules; generator and discriminator

* proposed `MultiModelDDP` class can use this way.
# trainer also has two nn.Modules, generator and discriminator;
# this way allocate DDP class to the each model in the trainer, not the trainer.
MultiModelDDP(trainer)

• Currently, pytorch lightning simple GAN training has has problem with DistributedDataParallel strategy. It tries to wrap pl.trainer, not the nn.Module models in the pl.trainer

• Although we can activate find_unused_parameters=True options to avoid this issue but it is not right way; I think it is just a trick.

• So the key idea to solve this issue is that we assign DistributedDataParallel to the each model in the pl.trainer, different from the previous strategy DDPStrategy.

• I already tested with my GPUs to visaulize the result and tracked the gradients of model with each epoch; it works and you can see the visulized result in thie google drive link

Before submitting

• Was this discussed/agreed via a GitHub issue? (not for typos and docs)
• [v] Did you read the contributor guideline, Pull Request section?
• [v] 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 minor internal changes/refactors)

PR review

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

Reviewer checklist

• 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

---

📚 Documentation preview 📚: https://pytorch-lightning--20936.org.readthedocs.build/en/20936/

[stale]

This pull request has been automatically marked as stale because it has not had recent activity. It will be closed in 7 days if no further activity occurs. If you need further help see our docs: https://lightning.ai/docs/pytorch/latest/generated/CONTRIBUTING.html#pull-request or ask the assistance of a core contributor here or on Discord. Thank you for your contributions.

[Borda]

@SkafteNicki could you pls check too :)

[Borda]

@samsara-ku please mark also the comments as resolved to keep track what has left :)

[SkafteNicki]

I think we are getting close with this PR, just a few suggestions

[samsara-ku]

I add some base test case for MultiModelDDPStrategy based on the contents of test_ddp.py as mentioned by @SkafteNicki .

I think I should add more test case about the new strategy like below:

1. Test for the "ONLY ABOUT GAN TRAINING" with this strategy; currently the base case mimicking the other test case didn't cover this kind of job not yet.

2. Something to check training convergence of training GAN; I don't have any idea for this, so I'll take the time for this.

[samsara-ku]

Hi @Borda, I’m picking this up again and have updated the PR with refined test coverage and compatibility fixes.

Could you share guidance on the expected test assertions (especially for GAN convergence/validation) and any remaining concerns with the current approach?

[samsara-ku]

I change my bad testcase codes in right way w/ MultiModelDDPStrategy class.

But, here is some problems I have so please anyone give me some right direction to solve this problem:

1. Need to figure out to check the training convergence of GAN

Currently the testcase check the training only with Discriminator's loss but it might be not the best way to validate the training

2. Fail to write test_find_unused_parameters_multi_model_ddp_raises codes

I have tried to take some time to test this case, but I cannot find this; it is hard to check the unused parameter because of that strategy