Clean up localsgd backup params

Author: H-Huang

torchft/local_sgd.py

@@ -10,7 +10,7 @@
 """
 import logging
 from types import TracebackType
-from typing import Any, Callable, Dict, Iterator, List, Mapping, Optional, Type
+from typing import Any, Callable, Dict, Iterator, List, Optional, Tuple, Type
 
 import torch
 from torch import nn, optim
@@ -59,8 +59,6 @@ def __init__(
         model: nn.Module,
         optimizer: optim.Optimizer,
         sync_every: int,
-        backup_device: Optional[torch.device] = None,
-        pin_memory: bool = True,
     ) -> None:
         """
         Args:
@@ -78,21 +76,8 @@ def __init__(
         self._local_step = 0
         self._sync_every = sync_every
         assert sync_every >= 1, "sync_every must be greater than or equal to 1"
-        device = backup_device or torch.device("cpu")
-        self._backup_parameters: Dict[str, torch.Tensor] = {}
-        for name, p in self._model.named_parameters():
-            t = torch.empty(*tuple(p.shape), dtype=p.dtype, device=device)
-            if (
-                pin_memory
-                and t.device == torch.device("cpu")
-                and torch.cuda.is_available()
-            ):
-                t = t.pin_memory()
-            self._backup_parameters[name] = t
 
         self._hooks: List[RemovableHandle] = []
-        # Need to copy the parameters to the host to be safe if we are on the first step.
-        self._save_parameters()
 
     def __enter__(self) -> "LocalSGD":
         # Add optimizer hook which increments the local step counter and syncs if necessary
@@ -108,37 +93,26 @@ def __exit__(
         traceback: Optional[TracebackType],
     ) -> bool:
         # Handle any cleanup or error handling here
-        if exc_type is not None:
-            # If an exception occurred, restore parameters
-            self._restore_parameters()
         # Clean up hooks
         for hook in self._hooks:
             hook.remove()
         self._hooks.clear()
 
         return False  # Propagate exceptions
 
-    def _save_parameters(self) -> None:
-        with torch.no_grad():
-            # TODO: consider running copy on a separate stream
-            for name, p in self._model.named_parameters():
-                self._backup_parameters[name].copy_(p.data, non_blocking=True)
-
-    def _restore_parameters(self) -> None:
-        with torch.no_grad():
-            # TODO: consider running copy on a separate stream
-            for name, p in self._model.named_parameters():
-                p.data.copy_(self._backup_parameters[name], non_blocking=False)
-
     def _step_post_hook(
-        self, _optim: optim.Optimizer, _args: List[object], _kwargs: Dict[str, object]
+        self, _optim: optim.Optimizer, _args: Tuple[Any, ...], _kwargs: Dict[str, Any]
     ) -> None:
         """
         This hook is registered on the optimizer and is called after the optimizer step.
         """
-        self._local_step += 1
-        if self._local_step >= self._sync_every:
-            self.sync()
+        try:
+            self._local_step += 1
+            if self._local_step >= self._sync_every:
+                self.sync()
+        except Exception as e:
+            self._manager.report_error(e)
+            raise
 
     def sync(self) -> None:
         """
@@ -151,15 +125,9 @@ def sync(self) -> None:
     def _perform_sync(self) -> None:
         """
         Performs the synchronization of the model weights across the manager.
-        This method is intended to be overridden by subclasses to implement custom
-        synchronization logic.
         """
-        self._average()
         if self._manager.should_commit():
-            self._save_parameters()
-        else:
-            # commit failed, restore from the backup parameters
-            self._restore_parameters()
+            self._average()
 
     def _average(self) -> None:
         # TODO: do we need to broadcast buffers like DDP does?
@@ -174,7 +142,7 @@ def _average(self) -> None:
             work.wait()
 
 
-class DiLoCo(LocalSGD):
+class DiLoCo:
     """
     DiLoCo is a subclass of LocalSGD that overrides the synchronization
     mechanism to average and synchronize the pseudogradients (delta of the previous global weight and current local weights).
@@ -197,21 +165,96 @@ def __init__(
                 "Using DiLoCo require synchronous quorum to be enabled. "
                 "Ensure that the manager is initialized with use_async_quorum=False"
             )
-        super().__init__(
-            manager, model, inner_optimizer, sync_every, backup_device, pin_memory
-        )
+        super().__init__()
+        self._manager = manager
+        self._model = model
+        self._local_optimizer = inner_optimizer
+        self._local_step = 0
+        self._sync_every = sync_every
+        assert sync_every >= 1, "sync_every must be greater than or equal to 1"
+
+        self._hooks: List[RemovableHandle] = []
         self._outer_optimizer = outer_optimizer
+        self._original_parameters: Dict[str, torch.Tensor] = {}
+        for name, p in self._model.named_parameters():
+            t = torch.empty(*tuple(p.shape), dtype=p.dtype, device=backup_device)
+            if (
+                pin_memory
+                and t.device == torch.device("cpu")
+                and torch.cuda.is_available()
+            ):
+                t = t.pin_memory()
+            self._original_parameters[name] = t
+
+        # Need to copy the parameters to the host to be safe if we are on the first step.
+        self._save_parameters()
+
+    def _save_parameters(self) -> None:
+        with torch.no_grad():
+            # TODO: consider running copy on a separate stream
+            for name, p in self._model.named_parameters():
+                self._original_parameters[name].copy_(p.data, non_blocking=True)
+
+    def _restore_parameters(self) -> None:
+        with torch.no_grad():
+            # TODO: consider running copy on a separate stream
+            for name, p in self._model.named_parameters():
+                p.data.copy_(self._original_parameters[name], non_blocking=False)
+
+    def __enter__(self) -> "DiLoCo":
+        # Add optimizer hook which increments the local step counter and syncs if necessary
+        self._hooks.append(
+            self._local_optimizer.register_step_post_hook(self._step_post_hook)
+        )
+        return self
+
+    def __exit__(
+        self,
+        exc_type: Optional[Type[BaseException]],
+        exc_value: Optional[BaseException],
+        traceback: Optional[TracebackType],
+    ) -> bool:
+        # Handle any cleanup or error handling here
+        # Clean up hooks
+        for hook in self._hooks:
+            hook.remove()
+        self._hooks.clear()
+
+        return False  # Propagate exceptions
+
+    def _step_post_hook(
+        self, _optim: optim.Optimizer, _args: Tuple[Any, ...], _kwargs: Dict[str, Any]
+    ) -> None:
+        """
+        This hook is registered on the optimizer and is called after the optimizer step.
+        """
+        try:
+            self._local_step += 1
+            if self._local_step >= self._sync_every:
+                self.sync()
+        except Exception as e:
+            self._manager.report_error(e)
+            raise
+
+    def sync(self) -> None:
+        """
+        Synchronizes and averages the model weights across the manager.
+        """
+        self._manager.start_quorum()
+        self._perform_sync()
+        self._local_step = 0
 
     def _perform_sync(self) -> None:
         """
         Overrides the sync method to calculate the pseugradient, average them across the manager group, and
         step using the outer optimizer.
         """
+        print("Performing DiLoCo sync", flush=True)
 
         # Set the .grad field of each parameter to its pseudogradient
         for name, p in self._model.named_parameters():
-            assert name in self._backup_parameters
-            pseudogradient = p.data - self._backup_parameters[name]
+            assert name in self._original_parameters
+            pseudogradient = p.data - self._original_parameters[name]
             p.grad = pseudogradient
 
         self._average_grads()