[1/2] Collaborative Strategy

[SeanNaren]

What does this PR do?

Related #12647

This PR represents the code portion of the strategy. The next PR will be the documentation.

Does your PR introduce any breaking changes? If yes, please list them.

None

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 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:

• 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 🙃

cc @Borda @akihironitta @justusschock

[awaelchli]

this is an awesome addition. can't wait for the full docs!

[SeanNaren]

Huge thanks @awaelchli @justusschock for going through and reviewing, really appreciate it ❤️

Two things I need to address:

• Maybe we should consider inheriting from the SingleDeviceStrategy instead as currently only single device is supported?
• Should we mark this strategy as experimental/stable?
  • Given most of this code is a wrapper over hivemind, I think calling it experimental is a bit overkill but I can be swayed either way

[justusschock]

Maybe we should consider inheriting from the SingleDeviceStrategy instead as currently only single device is supported?

IMO we shouldn't do that. It has one single device per process, but in total allows training on many devices. That way it behaves like an elastic multi-node DDP if the processes are spawned externally.

Should we mark this strategy as experimental/stable?
Given most of this code is a wrapper over hivemind, I think calling it experimental is a bit overkill but I can be swayed either way

I think, I'd still mark it as experimental. This gives us more freedom to change stuff (like the per-node DDP thingy you mentioned).
Also we can mark it as stable whenever we want later on.
I think marking this directly as stable may yield false expectations from users.

[ethanwharris]

LGTM, just some small comments 😃

[kaushikb11]

🔥

[rohitgr7]

still reviewing

[rohitgr7]

awesome 🚀

[Borda]

gatekeeper/pass: @SeanNaren