Add support for init_meta_context, materialize_module

[tchaton]

What does this PR do?

Fixes #9375

This PR builds on top of pytorch/pytorch#66317. A code section will be dropped once merged to PyTorch.

The goal is no code change for the end users.

from pytorch_lightning.utilities.meta import init_meta_context, materialize_module

class MLP(nn.Module):
    def __init__(self, num_convs: int):
        super().__init__()
        self.lins = []
        for _ in range(num_convs):
            self.lins.append(nn.Linear(1, 1))
        self.layer = nn.Sequential(*self.lins)

with init_meta_context():
    m = nn.Linear(in_features=1, out_features=1)
    assert m.weight.device.type == "meta"
    mlp = MLP(4)
    assert mlp.layer[0].weight.device.type == "meta"

    materialize_module(mlp)
    assert mlp.layer[0].weight.device.type == "cpu"

m = nn.Linear(in_features=1, out_features=1)
assert m.weight.device.type == "cpu"

with init_meta_context():
    m = nn.Linear(in_features=1, out_features=1)
    assert m.weight.device.type == "meta"

m = nn.Linear(in_features=1, out_features=1)
assert m.weight.device.type == "cpu"

TODO:

• Verify it works with DeepSpeed on BoringModel

    with init_meta_context():
        model = BoringModel()
    assert model.layer.weight.device.type == "meta"
    trainer = Trainer(
        default_root_dir=tmpdir, plugins=[DeepSpeedPlugin(stage=3)], gpus=2, fast_dev_run=True, precision=16
    )
    trainer.fit(model)
    assert model.layer.weight.device.type == "cpu"

• Verify it works with DeepSpeed on minGPT. Find the associated example here: https://github.com/SeanNaren/minGPT/blob/meta_device/train.py

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 🙃

[SeanNaren]

This is incredible work allowing users to not have to change their model definition, and hopefully in the future FSDP supports sharding directly from the meta device modules.

(cc @myleott @blefaudeux @anj-s who may have some stuff to say about the API/future integration!)

Should we wait for the code to be merged into PyTorch and available in the nightly build?

EDIT: looping @jeffra and @tjruwase from the DeepSpeed team as well :)

[tchaton]

This is incredible work allowing users to not have to change their model definition, and hopefully in the future FSDP supports sharding directly from the meta device modules.

(cc @myleott @blefaudeux @anj-s who may have some stuff to say about the API/future integration!)

Should we wait for the code to be merged into PyTorch and available in the nightly build?

Adding @cbalioglu to the conversation.

IMO, we definitely want this for Lightning v1.5 as it is already working with PyTorch 1.10 nightly.
Hopefully, this feature will be merged in 1.10 and released before Lightning v1.5, so we can remove the copy / paste.

Best,
T.C

[SeanNaren]

Lets chase up on @zou3519 comment before proceeding further with this PR! We shouldn't merge till we understand the edge cases that are causing unstability.

Also I think it would be beneficial if we showed a use case of benefit (which I think @tchaton has with DeepSpeed but might need more testing based on instantiation times vs configure_sharded_model) before considering to merge this PR!

[cbalioglu]

Looks good to me as an experimental API. Please consider me your point of contact for any customer feedback or issue specific to the parts you copied over from the PyTorch PR #66317.