Support sharded optimizer state dumping outside of sharded strategies

src/pytorch_lightning/CHANGELOG.md

@@ -12,7 +12,8 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Added prefix to log message in `seed_everything` with rank info ([#13290](https://github.com/Lightning-AI/lightning/issues/13290))
 
 
--
+- Added support for saving sharded optimizer state dict outside of `DDPShardedStrategy` ([#14208](https://github.com/PyTorchLightning/pytorch-lightning/pull/14208))
+
 
 
 ### Changed

src/pytorch_lightning/strategies/sharded.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from contextlib import contextmanager
-from typing import Dict, Generator, List, Optional, Tuple, Union
+from typing import Dict, Generator, List, Tuple, Union
 
 from torch import Tensor
 from torch.nn import Module
@@ -27,7 +27,6 @@
 from pytorch_lightning.utilities.exceptions import MisconfigurationException
 from pytorch_lightning.utilities.imports import _FAIRSCALE_AVAILABLE, _FAIRSCALE_OSS_FP16_BROADCAST_AVAILABLE
 from pytorch_lightning.utilities.optimizer import optimizers_to_device
-from pytorch_lightning.utilities.rank_zero import rank_zero_only
 
 if _FAIRSCALE_AVAILABLE:
     from fairscale.nn.data_parallel.sharded_ddp import ShardedDataParallel
@@ -120,20 +119,6 @@ def _reinit_optimizers_with_oss(self, optimizers: List[Union[Optimizer, Lightnin
                 del optimizer
         return optimizers
 
-    def optimizer_state(self, optimizer: "OSS") -> Optional[dict]:
-        if isinstance(optimizer, LightningOptimizer):
-            optimizer = optimizer._optimizer
-        optimizer.consolidate_state_dict()
-        return self._optim_state_dict(optimizer)
-
-    @rank_zero_only
-    def _optim_state_dict(self, optimizer):
-        """
-        Retrieves state dict only on rank 0, which contains the entire optimizer state after calling
-        :meth:`consolidate_state_dict`.
-        """
-        return optimizer.state_dict()
-
     def pre_backward(self, closure_loss: Tensor) -> None:
         pass
 

src/pytorch_lightning/strategies/sharded_spawn.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from contextlib import contextmanager
-from typing import Any, Dict, Generator, List, Tuple
+from typing import Dict, Generator, List, Tuple
 
 from torch import Tensor
 from torch.nn import Module
@@ -25,7 +25,6 @@
 from pytorch_lightning.utilities.exceptions import MisconfigurationException
 from pytorch_lightning.utilities.imports import _FAIRSCALE_AVAILABLE
 from pytorch_lightning.utilities.optimizer import optimizers_to_device
-from pytorch_lightning.utilities.rank_zero import rank_zero_only
 
 if _FAIRSCALE_AVAILABLE:
     from fairscale.nn.data_parallel.sharded_ddp import ShardedDataParallel
@@ -85,11 +84,6 @@ def _wrap_optimizers(self, optimizers: List[Optimizer]) -> List["OSS"]:
 
         return self._reinit_optimizers_with_oss(optimizers)
 
-    def optimizer_state(self, optimizer: "OSS") -> Dict[str, Any]:
-        if isinstance(optimizer, OSS):
-            optimizer.consolidate_state_dict()
-        return self._optim_state_dict(optimizer)
-
     @contextmanager
     def block_backward_sync(self) -> Generator:
         """Blocks syncing gradients behaviour on backwards pass.
@@ -103,14 +97,6 @@ def block_backward_sync(self) -> Generator:
         else:
             yield None
 
-    @rank_zero_only
-    def _optim_state_dict(self, optimizer: Optimizer) -> Dict[str, Any]:
-        """
-        Retrieves state dict only on rank 0, which contains the entire optimizer state after calling
-        :meth:`consolidate_state_dict`.
-        """
-        return optimizer.state_dict()
-
     def pre_backward(self, closure_loss: Tensor) -> None:
         pass
 

src/pytorch_lightning/strategies/strategy.py

@@ -170,6 +170,16 @@ def optimizer_state(self, optimizer: Optimizer) -> Dict[str, Tensor]:
 
         Allows for syncing/collating optimizer state from processes in custom plugins.
         """
+        if isinstance(optimizer, LightningOptimizer):
+            optimizer = optimizer._optimizer
+
+        if hasattr(optimizer, "consolidate_state_dict"):
+            # there are optimizers like Fairscale's OSS or PyTorch's ZeroRedundancyOptimizer that shard their
+            # states, and to avoid OOM we consolidate the full state on rank 0 only
+            optimizer.consolidate_state_dict()
+            return optimizer.state_dict() if self.is_global_zero else {}
+
+        # for optimizers that are not sharded, we return the state dict on all ranks
         return optimizer.state_dict()
 
     def backward(

tests/tests_pytorch/strategies/test_ddp_strategy.py

@@ -24,8 +24,14 @@
 from pytorch_lightning.plugins.environments import ClusterEnvironment, LightningEnvironment
 from pytorch_lightning.strategies import DDPStrategy
 from pytorch_lightning.trainer.states import TrainerFn
+from pytorch_lightning.utilities.imports import _FAIRSCALE_AVAILABLE, _TORCH_GREATER_EQUAL_1_10
 from tests_pytorch.helpers.runif import RunIf
 
+if _FAIRSCALE_AVAILABLE:
+    from fairscale.optim import OSS
+if _TORCH_GREATER_EQUAL_1_10:
+    from torch.distributed.optim import ZeroRedundancyOptimizer
+
 
 class BoringModelGPU(BoringModel):
     def on_train_start(self) -> None:
@@ -252,3 +258,50 @@ def test_ddp_strategy_set_timeout(mock_init_process_group):
     mock_init_process_group.assert_called_with(
         process_group_backend, rank=global_rank, world_size=world_size, timeout=test_timedelta
     )
+
+
+class BoringFairScaleOptimizerModel(BoringModel):
+    def configure_optimizers(self):
+        base_optimizer = torch.optim.SGD(self.layer.parameters(), lr=0.1)
+        return OSS(params=base_optimizer.param_groups, optim=type(base_optimizer), **base_optimizer.defaults)
+
+
+@RunIf(min_cuda_gpus=2, skip_windows=True, fairscale=True)
+@pytest.mark.parametrize("strategy", (pytest.param("ddp", marks=RunIf(standalone=True)), "ddp_spawn"))
+def test_ddp_strategy_checkpoint_multi_gpu_fairscale_optimizer(tmpdir, strategy):
+    """Test to ensure that checkpoint is saved correctly when using faircale optimizer."""
+    model = BoringFairScaleOptimizerModel()
+    trainer = Trainer(accelerator="gpu", devices=2, strategy=strategy, max_epochs=2)
+
+    trainer.fit(model)
+
+    checkpoint_path = os.path.join(tmpdir, "model.pt")
+    trainer.save_checkpoint(checkpoint_path)
+    saved_model = BoringModel.load_from_checkpoint(checkpoint_path)
+
+    # Assert model parameters are identical after loading
+    for ddp_param, shard_param in zip(model.parameters(), saved_model.parameters()):
+        assert torch.equal(ddp_param.to("cpu"), shard_param)
+
+
+class BoringZeroRedundancyOptimizerModel(BoringModel):
+    def configure_optimizers(self):
+        return ZeroRedundancyOptimizer(self.layer.parameters(), optimizer_class=torch.optim.Adam, lr=0.1)
+
+
+@RunIf(min_cuda_gpus=2, skip_windows=True, min_torch="1.10")
+@pytest.mark.parametrize("strategy", (pytest.param("ddp", marks=RunIf(standalone=True)), "ddp_spawn"))
+def test_ddp_strategy_checkpoint_zero_redundancy_optimizer(tmpdir, strategy):
+    """Test to ensure that checkpoint is saved correctly when using zero redundancy optimizer."""
+    model = BoringZeroRedundancyOptimizerModel()
+    trainer = Trainer(accelerator="gpu", devices=2, strategy=strategy, max_epochs=2)
+
+    trainer.fit(model)
+
+    checkpoint_path = os.path.join(tmpdir, "model.pt")
+    trainer.save_checkpoint(checkpoint_path)
+    saved_model = BoringModel.load_from_checkpoint(checkpoint_path)
+
+    # Assert model parameters are identical after loading
+    for ddp_param, shard_param in zip(model.parameters(), saved_model.parameters()):
+        assert torch.equal(ddp_param.to("cpu"), shard_param)

tests/tests_pytorch/strategies/test_sharded_strategy.py

@@ -14,6 +14,7 @@
 
 if _FAIRSCALE_AVAILABLE:
     from fairscale.nn.data_parallel.sharded_ddp import ShardedDataParallel
+    from fairscale.optim import OSS
 
 
 @pytest.mark.parametrize("clip_val", [0, 10])
@@ -314,3 +315,27 @@ def test_block_backward_sync():
 def test_ddp_kwargs_from_registry(strategy_name, expected_ddp_kwargs):
     trainer = Trainer(strategy=strategy_name)
     assert trainer.strategy._ddp_kwargs == expected_ddp_kwargs
+
+
+class BoringFairScaleOptimizerModel(BoringModel):
+    def configure_optimizers(self):
+        base_optimizer = torch.optim.SGD(self.layer.parameters(), lr=0.1)
+        return OSS(params=base_optimizer.param_groups, optim=type(base_optimizer), **base_optimizer.defaults)
+
+
+@RunIf(min_cuda_gpus=2, skip_windows=True, fairscale=True)
+@pytest.mark.parametrize("strategy", (pytest.param("ddp_sharded", marks=RunIf(standalone=True)), "ddp_sharded_spawn"))
+def test_ddp_sharded_strategy_checkpoint_multi_gpu_fairscale_optimizer(tmpdir, strategy):
+    """Test to ensure that checkpoint is saved correctly when using fairscale optimizers."""
+    model = BoringFairScaleOptimizerModel()
+    trainer = Trainer(accelerator="gpu", devices=2, strategy=strategy, max_epochs=2)
+
+    trainer.fit(model)
+
+    checkpoint_path = os.path.join(tmpdir, "model.pt")
+    trainer.save_checkpoint(checkpoint_path)
+    saved_model = BoringModel.load_from_checkpoint(checkpoint_path)
+
+    # Assert model parameters are identical after loading
+    for ddp_param, shard_param in zip(model.parameters(), saved_model.parameters()):
+        assert torch.equal(ddp_param.to("cpu"), shard_param)