Add ShardedTensor support in LightningModule

[yifuwang]

Register ShardedTensor state dict hooks in LightningModule.__init__
if the pytorch version supports ShardedTensor.

What does this PR do?

Register ShardedTensor state dict hooks in LightningModule.__init__ if the pytorch version supports ShardedTensor (#8633).

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 #8944 (447af29) into master (c8e9fb4) will decrease coverage by 4%.
The diff coverage is 62%.

@@           Coverage Diff           @@
##           master   #8944    +/-   ##
=======================================
- Coverage      93%     89%    -4%     
=======================================
  Files         175     175            
  Lines       14457   14465     +8     
=======================================
- Hits        13406   12811   -595     
- Misses       1051    1654   +603     

[ananthsub]

FYI @SeanNaren @tchaton

[justusschock]

@yifuwang @ananthsub do we need to do the same to the metrics modules in torchmetrics?

[SeanNaren]

Thanks for this, @yifuwang @ananthsub in the future would you see an API like this where we can handle distributed communications?

class BoringModelWithShardedTensor(BoringModel):
    def __init__(self, spec):
        super().__init__()
        self.spec = spec

    def configure_sharded_model(self):
        self.sharded_tensor = dist._sharded_tensor.empty(self.spec, 10, 20)
        self.sharded_tensor.local_shards()[0].tensor.fill_(0)

model = BoringModelWithShardedTensor()
trainer = Trainer()
trainer.fit(model)

[yifuwang]

@yifuwang @ananthsub do we need to do the same to the metrics modules in torchmetrics?

Good question. I'm not aware of any metric implementation that maintains sharded state. If there is, they should ideally be represented with sharded tensor once the feature's released, in which case this change also applies in torchmetrics.

[yifuwang]

Thanks for this, @yifuwang @ananthsub in the future would you see an API like this where we can handle distributed communications?

class BoringModelWithShardedTensor(BoringModel):
    def __init__(self, spec):
        super().__init__()
        self.spec = spec

    def configure_sharded_model(self):
        self.sharded_tensor = dist._sharded_tensor.empty(self.spec, 10, 20)
        self.sharded_tensor.local_shards()[0].tensor.fill_(0)

model = BoringModelWithShardedTensor()
trainer = Trainer()
trainer.fit(model)

Very good question. I think it will be inevitable for LightningModule to recommend a best practice for organizing model parallel logic. However it's still not clear to me what it should look like.

I think the API above makes a lot of sense for model parallel uses cases that are developed on Lightning. However, for a self-contained model parallel use cases that just want to train with Lightning, there's a chance that the hook may be too opinionated.

It's definitely difficult to strike a good balance between usability and flexibility. I think it'd be helpful for us to do a informal survey of model parallel use cases we'd like to support in order to make an informed decision.

[s-rog]

This is an edge case, but the version check passes for 1.10.0a0+ecc3718 (the pytorch version in NGC 21.07) but the subsequent import fails.

This is probably the result of Nvidia's cherry picking, and I'm unsure if this is something we want to support?