Feature add ZeRO support to Checkpoint in a distributed configuration

ignite/handlers/checkpoint.py

@@ -11,6 +11,14 @@
 
 import torch
 import torch.nn as nn
+from packaging.version import Version
+
+if Version(torch.__version__) >= Version("1.9.0"):
+    from torch.distributed.optim import ZeroRedundancyOptimizer
+
+    HAVE_ZERO = True
+else:
+    HAVE_ZERO = False
 
 import ignite.distributed as idist
 from ignite.base import Serializable
@@ -166,13 +174,14 @@ class Checkpoint(Serializable):
             > checkpoint_12345.pt
 
     Note:
-        This class is distributed configuration-friendly: it is not required to instantiate the class in rank 0
-        process only.
+        This class is distributed configuration-friendly: it is not required to instantiate the class in rank 0 only
+        process. This class supports automatically distributed configuration and if used with
+        :class:`~ignite.handlers.DiskSaver`, checkpoint is stored by rank 0 process.
 
     .. warning::
 
-        When running on XLA devices, it should be run in all processes, otherwise application can get stuck on
-        saving the checkpoint.
+        When running on XLA devices or using :class:`~torch.distributed.optim.ZeroRedundancyOptimizer`, it
+        should be run in all processes, otherwise application can get stuck while saving the checkpoint.
 
         .. code-block:: python
 
@@ -282,7 +291,7 @@ def __init__(
         filename_pattern: Optional[str] = None,
         include_self: bool = False,
         greater_or_equal: bool = False,
-        save_on_rank: Optional[int] = 0,
+        save_on_rank: int = 0,
     ):
 
         if not isinstance(to_save, collections.Mapping):
@@ -466,6 +475,10 @@ def _setup_checkpoint(self) -> Dict[str, Dict[Any, Any]]:
             for k, obj in self.to_save.items():
                 if isinstance(obj, (nn.DataParallel, nn.parallel.DistributedDataParallel)):
                     obj = obj.module
+                elif HAVE_ZERO and isinstance(obj, ZeroRedundancyOptimizer):
+                    obj.consolidate_state_dict(to=self.save_on_rank)
+                    if self.save_on_rank != idist.get_rank():
+                        continue
                 checkpoint[k] = obj.state_dict()
         return checkpoint
 
@@ -782,7 +795,7 @@ def __init__(
         atomic: bool = True,
         create_dir: bool = True,
         require_empty: bool = True,
-        save_on_rank: Optional[int] = 0,
+        save_on_rank: int = 0,
         **kwargs: Any,
     ):
         self.dirname = Path(dirname).expanduser()
@@ -948,7 +961,7 @@ def __init__(
         filename_pattern: Optional[str] = None,
         include_self: bool = False,
         greater_or_equal: bool = False,
-        save_on_rank: Optional[int] = 0,
+        save_on_rank: int = 0,
         **kwargs: Any,
     ):
 

tests/ignite/handlers/test_checkpoint.py

@@ -1243,9 +1243,37 @@ def _test_checkpoint_load_objects_ddp(device):
     Checkpoint.load_objects(to_load, checkpoint)
 
 
+def _test_checkpoint_with_ZeRO(device, dirname, local_rank):
+
+    from torch.distributed.optim import ZeroRedundancyOptimizer
+
+    model = DummyModel().to(device)
+    opt = ZeroRedundancyOptimizer(model.parameters(), torch.optim.SGD, lr=0.01)
+    mocked_opt = MagicMock(ZeroRedundancyOptimizer, wraps=opt)
+
+    # A `step` should be called to optimizer state get populated.
+    out = model(torch.Tensor([1.0]))
+    out.backward()
+    mocked_opt.step()
+
+    to_save = {"model": model, "optim": mocked_opt}
+    checkpointer = Checkpoint(to_save, dirname, save_on_rank=1)
+
+    engine = Engine(lambda e, b: None)
+    checkpointer(engine)
+
+    mocked_opt.consolidate_state_dict.assert_called_once_with(to=1)
+
+    if local_rank == 1:
+
+        loaded_state_dict = torch.load(dirname / "checkpoint_0.pt", map_location=device)["optim"]
+        state_dict = opt.state_dict()
+        assert loaded_state_dict == state_dict
+
+
 @pytest.mark.distributed
 @pytest.mark.skipif(not idist.has_native_dist_support, reason="Skip if no native dist support")
-def test_distrib_gloo_cpu_or_gpu(distributed_context_single_node_gloo, get_rank_zero_dirname):
+def test_distrib_gloo_cpu_or_gpu(distributed_context_single_node_gloo, dirname, get_rank_zero_dirname, local_rank):
 
     device = idist.device()
     rank_zero_dirname = get_rank_zero_dirname()
@@ -1254,6 +1282,11 @@ def test_distrib_gloo_cpu_or_gpu(distributed_context_single_node_gloo, get_rank_
     _test_checkpoint_with_ddp(device)
     _test_checkpoint_load_objects_ddp(device)
 
+    from ignite.handlers.checkpoint import HAVE_ZERO
+
+    if HAVE_ZERO:
+        _test_checkpoint_with_ZeRO(device, dirname, local_rank)
+
 
 @pytest.mark.distributed
 @pytest.mark.skipif(not idist.has_native_dist_support, reason="Skip if no native dist support")