Fix Adagrad optimizer not working with DDP/GPU (#7277)

Co-authored-by: ananthsub <ananth.subramaniam@gmail.com>
Co-authored-by: thomas chaton <thomas@grid.ai>

CHANGELOG.md

@@ -440,6 +440,9 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Fixed `apex` not properly instantiated when running with `ddp` ([#7274](https://github.com/PyTorchLightning/pytorch-lightning/pull/7274))
 
 
+- Fixed optimizer `state` not moved to `GPU` ([#7277](https://github.com/PyTorchLightning/pytorch-lightning/pull/7277))
+
+
 ## [1.2.7] - 2021-04-06
 
 ### Fixed

pytorch_lightning/accelerators/accelerator.py

@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import contextlib
+from collections import defaultdict
 from typing import Any, Callable, Dict, Generator, Iterable, List, Optional, Union
 
 import torch
@@ -25,7 +26,7 @@
 from pytorch_lightning.plugins.training_type import TrainingTypePlugin
 from pytorch_lightning.trainer.states import TrainerState
 from pytorch_lightning.utilities import _NATIVE_AMP_AVAILABLE, rank_zero_warn
-from pytorch_lightning.utilities.apply_func import move_data_to_device
+from pytorch_lightning.utilities.apply_func import apply_to_collection, move_data_to_device
 from pytorch_lightning.utilities.enums import AMPType, GradClipAlgorithmType, LightningEnum
 from pytorch_lightning.utilities.types import EPOCH_OUTPUT, STEP_OUTPUT
 
@@ -102,11 +103,22 @@ def start_predicting(self, trainer: 'pl.Trainer') -> None:
 
     def pre_dispatch(self, trainer: 'pl.Trainer') -> None:
         """Hook to do something before the training/evaluation/prediction starts."""
+        self._move_optimizer_state()
+
         self.training_type_plugin.pre_dispatch()
         if self.training_type_plugin.setup_optimizers_in_pre_dispatch:
             self.setup_optimizers(trainer)
+
         self.precision_plugin.pre_dispatch()
 
+    def _move_optimizer_state(self) -> None:
+        """ Moves the state of the optimizers to the GPU if needed. """
+        for opt in self.optimizers:
+            state = defaultdict(dict)
+            for p, v in opt.state.items():
+                state[p] = apply_to_collection(v, torch.Tensor, move_data_to_device, self.root_device)
+            opt.state = state
+
     def dispatch(self, trainer: 'pl.Trainer') -> None:
         """Hook to do something before the training/evaluation/prediction starts."""
         self.training_type_plugin.dispatch(trainer)

tests/trainer/optimization/test_optimizers.py

@@ -19,6 +19,7 @@
 from pytorch_lightning.utilities.exceptions import MisconfigurationException
 from tests.base import EvalModelTemplate
 from tests.helpers.boring_model import BoringModel
+from tests.helpers.runif import RunIf
 
 
 def test_optimizer_with_scheduling(tmpdir):
@@ -498,3 +499,28 @@ def configure_optimizers(self):
     trainer = Trainer(default_root_dir=tmpdir, fast_dev_run=True)
     with pytest.warns(RuntimeWarning, match='the keys will be ignored'):
         trainer.fit(model)
+
+
+class TestModel(BoringModel):
+
+    def configure_optimizers(self):
+        # Adagrad creates state tensors immediately, model is not yet on GPU.
+        return torch.optim.Adagrad(self.parameters())
+
+    def on_train_start(self, *args, **kwargs):
+        opt = self.optimizers()
+        _, state = next(iter(opt.state.items()))
+        assert state["sum"].device == torch.device("cuda", self.local_rank) == self.device
+
+
+@RunIf(min_gpus=2, special=True)
+def test_optimizer_state_on_device(tmpdir):
+    """ Test that optimizers that create state initially at instantiation still end up with the state on the GPU. """
+    model = TestModel()
+    trainer = Trainer(
+        default_root_dir=tmpdir,
+        gpus=2,
+        accelerator="ddp",
+        fast_dev_run=True,
+    )
+    trainer.fit(model)