Rename managed_* to externally_managed_*

These functions rely on the caller keeping track of which parameters
have been trained, and passing that information in the function call.

mammoth/distributed/__init__.py

@@ -4,7 +4,7 @@
     batch_producer,
     consumer,
     broadcast_tensors,
-    managed_reduce_and_rescale_grads,
+    externally_managed_reduce_and_rescale_grads,
     ErrorHandler,
 )
 from mammoth.distributed.contexts import DeviceContext, WorldContext, DeviceContextEnum
@@ -20,7 +20,7 @@
     "batch_producer",
     "broadcast_tensors",
     "consumer",
-    "managed_reduce_and_rescale_grads",
+    "externally_managed_reduce_and_rescale_grads",
     "ErrorHandler",
     "DeviceContext",
     "WorldContext",

mammoth/distributed/communication.py

@@ -36,7 +36,7 @@ def broadcast_tensors(tensors, src=0, group=None):
             torch.distributed.broadcast(t, src, group=group)
 
 
-def managed_reduce_and_rescale_grads(
+def externally_managed_reduce_and_rescale_grads(
     named_parameters,
     has_local_gradient: bool,
     gradient_norm: int,

mammoth/trainer.py

@@ -278,7 +278,7 @@ def train(
                     continue
                 # logger.warning(f'Syncing {component.get_name()}')   # DEBUG
                 params = component.named_parameters(self.model)
-                mammoth.distributed.managed_reduce_and_rescale_grads(
+                mammoth.distributed.externally_managed_reduce_and_rescale_grads(
                     named_parameters=params,
                     has_local_gradient=gradient_sync.has_local_gradient,
                     gradient_norm=gradient_sync.gradient_norm,
@@ -288,7 +288,7 @@ def train(
             self._maybe_update_stats_from_parameters(report_stats, self.model.named_parameters())
 
             # Including single-device components
-            self.optim.managed_step(gradient_syncs)
+            self.optim.externally_managed_step(gradient_syncs)
             self.optim.zero_grad()
 
             if step % 1000 == 0 and step > 0:

mammoth/utils/optimizers.py

@@ -202,7 +202,7 @@ def zero_grad(self):
         for name in self.optimizers:
             self.optimizers[name].zero_grad()
 
-    def managed_step(self, gradient_syncs, grad_scaler=None):
+    def externally_managed_step(self, gradient_syncs, grad_scaler=None):
         """Step through only the trained suboptimizers"""
         trained_components = {
             gradient_sync.component.get_name() for gradient_sync in gradient_syncs
@@ -387,7 +387,7 @@ def backward(self, loss):
         else:
             loss.backward()
 
-    def managed_step(self, *args, **kwargs):
+    def externally_managed_step(self, *args, **kwargs):
         """Update the model parameters based on current gradients.
 
         Optionally, will employ gradient modification or update learning
@@ -415,7 +415,7 @@ def managed_step(self, *args, **kwargs):
             # Updates the scale for next iteration.
             self._scaler.update()
         else:
-            self._optimizer.managed_step(*args, **kwargs)
+            self._optimizer.externally_managed_step(*args, **kwargs)
         self._decay_step += 1
         self._training_step += 1