Disable strict loading in multiprocessing launcher

[awaelchli]

What does this PR do?

Fixes #14534

Sets strict=False when loading the state dict of the model back into the main process. The model in the main process may have a different architecture than the one trained in the worker processes:

class BoringModel(LightningModule):
    def __init__(self):
        super().__init__()
        self.layer1 = torch.nn.Linear(32, 2)

    def setup(self, stage=None):
        self.layer2 = torch.nn.Linear(32, 2)  # this layer does not exist in main process


model = BoringModel()  # layer2 does not exist yet
trainer = Trainer(strategy="ddp_spawn", ...)
trainer.fit(model) 
# here, at the end of fit, we load model weights back into main process
# but we can only do so for layer1, because layer2 does not exist in main process

This is a limitation of this type of training with the "spawn" method. Since we don't know what the user will do with the model after fit(), the best we can do is load the weights that match.

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 update the CHANGELOG? (not for typos, docs, test updates, or internal minor changes/refactorings)

PR review

Anyone in the community is free to review the PR once the tests have passed.
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?

I made sure I had fun coding 🙃

cc @Borda @justusschock @awaelchli

[github-actions]

⚡ Required checks status: All passing 🟢

Groups summary

🟢 pytorch_lightning: Tests workflow

Check ID	Status
pl-cpu (macOS-11, pytorch, 3.8, 1.11)	success	✅
pl-cpu (macOS-11, pytorch, 3.9, 1.12)	success	✅
pl-cpu (macOS-11, pytorch, 3.10, 1.13)	success	✅
pl-cpu (macOS-11, pytorch, 3.8, 1.10, oldest)	success	✅
pl-cpu (ubuntu-20.04, pytorch, 3.8, 1.10)	success	✅
pl-cpu (ubuntu-20.04, pytorch, 3.9, 1.11)	success	✅
pl-cpu (ubuntu-20.04, pytorch, 3.10, 1.12)	success	✅
pl-cpu (ubuntu-20.04, pytorch, 3.10, 1.13)	success	✅
pl-cpu (ubuntu-20.04, pytorch, 3.7, 1.10, oldest)	success	✅
pl-cpu (windows-2022, pytorch, 3.9, 1.11)	success	✅
pl-cpu (windows-2022, pytorch, 3.10, 1.12)	success	✅
pl-cpu (windows-2022, pytorch, 3.10, 1.13)	success	✅
pl-cpu (windows-2022, pytorch, 3.7, 1.10, oldest)	success	✅
pl-cpu (slow, macOS-11, pytorch, 3.7, 1.11)	success	✅
pl-cpu (slow, ubuntu-20.04, pytorch, 3.7, 1.11)	success	✅
pl-cpu (slow, windows-2022, pytorch, 3.7, 1.11)	success	✅
pl-cpu (macOS-11, lightning, 3.8, 1.13)	success	✅
pl-cpu (ubuntu-20.04, lightning, 3.8, 1.13)	success	✅
pl-cpu (windows-2022, lightning, 3.8, 1.13)	success	✅

These checks are required after the changes to src/pytorch_lightning/strategies/launchers/multiprocessing.py, tests/tests_pytorch/strategies/launchers/test_multiprocessing.py, tests/tests_pytorch/strategies/test_ddp_spawn_strategy.py.

🟢 pytorch_lightning: Azure GPU

Check ID	Status
pytorch-lightning (GPUs)	success	✅

These checks are required after the changes to src/pytorch_lightning/strategies/launchers/multiprocessing.py, tests/tests_pytorch/strategies/launchers/test_multiprocessing.py, tests/tests_pytorch/strategies/test_ddp_spawn_strategy.py.

🟢 pytorch_lightning: Azure HPU

Check ID	Status
pytorch-lightning (HPUs)	success	✅

These checks are required after the changes to src/pytorch_lightning/strategies/launchers/multiprocessing.py, tests/tests_pytorch/strategies/launchers/test_multiprocessing.py, tests/tests_pytorch/strategies/test_ddp_spawn_strategy.py.

🟢 pytorch_lightning: Azure IPU

Check ID	Status
pytorch-lightning (IPUs)	success	✅

These checks are required after the changes to src/pytorch_lightning/strategies/launchers/multiprocessing.py, tests/tests_pytorch/strategies/launchers/test_multiprocessing.py, tests/tests_pytorch/strategies/test_ddp_spawn_strategy.py.

🟢 pytorch_lightning: Docs

Check ID	Status
make-doctest (pytorch)	success	✅
make-html (pytorch)	success	✅

These checks are required after the changes to src/pytorch_lightning/strategies/launchers/multiprocessing.py.

🟢 mypy

Check ID	Status
mypy	success	✅

These checks are required after the changes to src/pytorch_lightning/strategies/launchers/multiprocessing.py.

🟢 install

Check ID	Status
install-pkg (ubuntu-22.04, app, 3.7)	success	✅
install-pkg (ubuntu-22.04, app, 3.10)	success	✅
install-pkg (ubuntu-22.04, fabric, 3.7)	success	✅
install-pkg (ubuntu-22.04, fabric, 3.10)	success	✅
install-pkg (ubuntu-22.04, pytorch, 3.7)	success	✅
install-pkg (ubuntu-22.04, pytorch, 3.10)	success	✅
install-pkg (ubuntu-22.04, lightning, 3.7)	success	✅
install-pkg (ubuntu-22.04, lightning, 3.10)	success	✅
install-pkg (ubuntu-22.04, notset, 3.7)	success	✅
install-pkg (ubuntu-22.04, notset, 3.10)	success	✅
install-pkg (macOS-12, app, 3.7)	success	✅
install-pkg (macOS-12, app, 3.10)	success	✅
install-pkg (macOS-12, fabric, 3.7)	success	✅
install-pkg (macOS-12, fabric, 3.10)	success	✅
install-pkg (macOS-12, pytorch, 3.7)	success	✅
install-pkg (macOS-12, pytorch, 3.10)	success	✅
install-pkg (macOS-12, lightning, 3.7)	success	✅
install-pkg (macOS-12, lightning, 3.10)	success	✅
install-pkg (macOS-12, notset, 3.7)	success	✅
install-pkg (macOS-12, notset, 3.10)	success	✅
install-pkg (windows-2022, app, 3.7)	success	✅
install-pkg (windows-2022, app, 3.10)	success	✅
install-pkg (windows-2022, fabric, 3.7)	success	✅
install-pkg (windows-2022, fabric, 3.10)	success	✅
install-pkg (windows-2022, pytorch, 3.7)	success	✅
install-pkg (windows-2022, pytorch, 3.10)	success	✅
install-pkg (windows-2022, lightning, 3.7)	success	✅
install-pkg (windows-2022, lightning, 3.10)	success	✅
install-pkg (windows-2022, notset, 3.7)	success	✅
install-pkg (windows-2022, notset, 3.10)	success	✅

These checks are required after the changes to src/pytorch_lightning/strategies/launchers/multiprocessing.py.

---

Thank you for your contribution! 💜

Note
This comment is automatically generated and updates for 60 minutes every 180 seconds. If you have any other questions, contact carmocca for help.

[vitkl]

@awaelchli Thank you for quickly implementing this!

It is possible to install pytorch-lightning with this fix (not lightning)? I don't see any way to do this listed here https://pytorch-lightning.readthedocs.io/en/stable/starter/installation.html

[carmocca]

Try this:

PACKAGE_NAME=pytorch pip install https://github.com/Lightning-AI/lightning/archive/refs/heads/master.zip -U

[vitkl]

This works, thank you!

[vitkl]

Although the solution itself doesn't work for my pyro model - the module ends up with no parameters.

module # nn.Module that contain pyro model and pyro guide as attributes module.model and module.guide
training_plan = TrainingPlan(pyro_module=module, **plan_kwargs) # pl.LightningModule
trainer = Trainer(
    max_epochs=max_epochs,
    accelerator=accelerator,
    devices=devices,
    strategy=strategy,
    **trainer_kwargs,
)
trainer.fit(training_plan, data_splitter)
module.state_dict().keys()
# no parameters listed

Is there any point downstream of this change where the contents of state_dict can be ignored or overwritten?

[awaelchli]

This is probably because you don't create your layers at the time of instantiation, only later. You will always run into this limitation with the "ddp_spawn" strategy, it's a result of the design. In this case, you should choose strategy="ddp" and you will never have this issue.

[vitkl]

I see, thanks for explaining!

And there is no way to modify the strategy - e.g. run Callback setup before loading the parameters in the main process?

[awaelchli]

If you want, you can always call the setup() method yourself in the main process:

my_callback.setup("fit") # call so that layers exist after fit
my_model.setup("fit")  # call so that layers exist after fit
trainer.fit(my_model, ...)

But before falling back to this workaround, I suggest just using the regular ddp strategy.

[vitkl]

Thanks @awaelchli! Both solutions ddp and dpp_notebook + running callback in the main process seem to work (code runs). However, I see that both solutions lead to identical batches loaded to both devices (the data-loaded tensors such as the pyro plate indices on both devices are identical). I read this #7186 and other related issues - but I don't understand why this should be happening. Is it possible to check if DistributedSampler (replace_sampler_ddp=True) was created successfully?

It appears that worker_init_fn=pl_worker_init_function leads to all workers being initialised in all processes using the same seed. Is this expected?

[awaelchli]

Is it possible to check if DistributedSampler (replace_sampler_ddp=True) was created successfully?

You can check isinstance(trainer.train_dataloaders[0].sampler, DistributedSampler) for example.

It appears that worker_init_fn=pl_worker_init_function leads to all workers being initialised in all processes using the same seed. Is this expected?

Call seed_everything(1, workers=True) to seed dataloader workers based on global rank. This doesn't concern the training processes. You can set a different seed like so: seed_everything(seed + trainer.global_rank) for example.

For further questions, please consider posting in the forum or if you find a bug, a new issue would be appreciated (since this topic here is about strict loading of weights).