fix: model.set_requires_gradient_sync(False) should be called to turn off gradient synchronization in FSDP2 (#3762)

* fix :`model.set_requires_gradient_sync(False)` should be called to turn off gradient synchronization in FSDP2.

* fix: remove trailing whitespace

Author: EquationWalker

src/accelerate/accelerator.py

@@ -1169,13 +1169,20 @@ def no_sync(self, model):
         >>> optimizer.zero_grad()
         ```
         """
-        context = contextlib.nullcontext
-        if self.use_distributed:
-            if self.distributed_type != DistributedType.DEEPSPEED or self.state.deepspeed_plugin.zero_stage < 2:
-                context = getattr(model, "no_sync", context)
+        if self.is_fsdp2:
+            model.set_requires_gradient_sync(False)
+            try:
+                yield
+            finally:
+                model.set_requires_gradient_sync(True)
+        else:
+            context = contextlib.nullcontext
+            if self.use_distributed:
+                if self.distributed_type != DistributedType.DEEPSPEED or self.state.deepspeed_plugin.zero_stage < 2:
+                    context = getattr(model, "no_sync", context)
 
-        with context():
-            yield
+            with context():
+                yield
 
     @staticmethod
     @contextmanager