Add Deepspeed Zero 3 MiCS support (Issues #20378)

[hehepig4]

What does this PR do?

After deepspeed 0.9.2, they provided Mics support, which can specify how to split parameters across devices in Zero stage 3 reference.

To activate MiCS, users should add 'mics_shard_size' in deepspeed config, and use 'deepspeed.zero.MiCS_Init' instead of 'deepspeed.zero.Init' when initializing models. Issues occur when:

1. Specify mics_shard_size>0 while using deepspeed.zero.Init causes exceptions.

The core changes are done by altering zero.Init in DeepSpeedStrategy.model_sharded_context (line 519 in lightning.pytorch.strategies.deepspeed.py) when detecting 'mics_shard_size' in deepspeed zero_optmization config.

Fixes & adds support for #20378

Before submitting

• Was this discussed/agreed 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 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--20461.org.readthedocs.build/en/20461/

[codecov]

Codecov Report

Attention: Patch coverage is 0% with 10 lines in your changes missing coverage. Please review.

Project coverage is 88%. Comparing base (60289d7) to head (87eefdc).

Additional details and impacted files

@@           Coverage Diff           @@
##           master   #20461   +/-   ##
=======================================
- Coverage      88%      88%   -0%     
=======================================
  Files         267      267           
  Lines       23382    23389    +7     
=======================================
  Hits        20479    20479           
- Misses       2903     2910    +7     

[lantiga]

This is great, thanks for the contribution @hehepig4

[lantiga]

let's see how CI goes and we can proceed

[lantiga]

In the meantime @hehepig4, would you be willing to add a mention to this in the relevant docs section?

https://github.com/Lightning-AI/pytorch-lightning/blob/master/docs/source-pytorch/advanced/model_parallel/deepspeed.rst

Also ideally we could make the same exact change in Fabric as well:
https://github.com/Lightning-AI/pytorch-lightning/blob/master/src/lightning/fabric/strategies/deepspeed.py#L376

[hehepig4]

Thanks! I will first work on docs, then fabric.

[lantiga]

Looks great, thanks for the updates. I see CI is going out of memory, I need to further investigate.