wraps DDP models with DSD (#857)

Summary:
Pull Request resolved: #857

Distributed State Dict is the current suggested way from PyTorch for ensuring parallelized models state dicts are compatible with save/loads in Single process or re-sharding scenarios.

This diff updates dcp_saver to use DSD for DDP models. A good idea would be wrap all models in TNT with DSD, as this could replace some of the wrapper logic for FSDP and would guarantee future compat.

N5551629 also contains a workaround for current DDP model saved before this diff, by manually removing the "module." prefix in the checkpoint.

Differential Revision: D59234083

Author: LucasLLC

torchtnt/framework/callbacks/dcp_saver.py

@@ -20,7 +20,7 @@
 )
 from torch.distributed.checkpoint.planner import LoadPlanner, SavePlanner
 from torch.distributed.checkpoint.storage import StorageReader, StorageWriter
-
+from torch.nn.parallel import DistributedDataParallel
 from torchtnt.framework.callbacks._checkpoint_utils import (
     _prepare_app_state_for_checkpoint,
     _prepare_app_state_for_restore,
@@ -41,6 +41,7 @@
 from torchtnt.utils.checkpoint import BestCheckpointConfig, CheckpointPath
 from torchtnt.utils.optimizer import init_optim_state
 from torchtnt.utils.rank_zero_log import rank_zero_info, rank_zero_warn
+
 from torchtnt.utils.stateful import MultiStateful, Stateful
 
 
@@ -62,6 +63,48 @@
         FileSystemWriter as Writer,
     )
 
+# below code provides BC for PyTorch versions which don't include distributed state dict
+# TODO: remove below code once this path is not longer supported
+try:
+    import torch.distributed.checkpoint.state_dict as dsd
+
+    # pyre-ignore Incompatible variable type [9]
+    get_model_state_dict = dsd.get_model_state_dict
+
+    def set_model_state_dict(mod: torch.nn.Module, state_dict: Dict[str, Any]) -> None:
+        return dsd.set_model_state_dict(mod, state_dict)
+
+except ImportError:
+    logger.warn(
+        "torch.distributed.checkpoint.state_dict checkpoint is not available, "
+        "falling back on defaults. Consider updating PyTorch, as this version "
+        "will not be supported in the future."
+    )
+
+    def get_model_state_dict(mod: torch.nn.Module) -> Dict[str, Any]:
+        return mod.state_dict()
+
+    def set_model_state_dict(mod: torch.nn.Module, state_dict: Dict[str, Any]) -> None:
+        return mod.load_state_dict(state_dict)
+
+
+class DSDModelWrapper(Stateful):
+    """This wrapper converts state dicts to Distributed State Dicts, essentially generating
+    state dicts as if they were created using single-device methods. This is useful for
+    when checkpoint models might be resharded, or loaded in notebooks or otherwise non-distributed
+    settings.
+
+    """
+
+    def __init__(self, mod: torch.nn.Module) -> None:
+        self.mod: torch.nn.Module = mod
+
+    def state_dict(self) -> Dict[str, Any]:
+        return get_model_state_dict(self.mod)
+
+    def load_state_dict(self, state_dict: Dict[str, Any]) -> None:
+        set_model_state_dict(self.mod, state_dict)
+
 
 class DistributedCheckpointSaver(BaseCheckpointer):
     """
@@ -148,6 +191,11 @@ def _checkpoint_impl(
         curr_snapshot_wait = hook == "on_train_end"
 
         app_state = _prepare_app_state_for_checkpoint(state, unit, intra_epoch)
+
+        for key, obj in app_state.items():
+            if isinstance(obj, DistributedDataParallel):
+                app_state[key] = DSDModelWrapper(obj)
+
         # TODO: evaluate whether we need to implement the equivalent of torchsnapshot.RNGState()
         if self._async_checkpoint:
             with get_timing_context(state, f"{self.__class__.__name__}.async_save"):
@@ -315,14 +363,17 @@ def restore(
                     )
 
         # necessary for loading optimizers since states are initialized lazy
-        for obj in app_state.values():
+        for key, obj in app_state.items():
             # sometimes optimizers are actually held in a wrapper which handles calling
             # state_dict and load_state_dict, sa is the case for
             # `torchtnt.utils.prepare_module.FSDPOptimizerWrapper`, this handles that case.
             optimizer = getattr(obj, "optimizer", obj)
             if isinstance(optimizer, torch.optim.Optimizer):
                 init_optim_state(optimizer)
 
+            if isinstance(obj, DistributedDataParallel):
+                app_state[key] = DSDModelWrapper(obj)
+
         try:
             dcp.load(
                 {"app_state": MultiStateful(app_state)},