fix call to manager allreduce (meta-pytorch#273)

Summary:
Pull Request resolved: meta-pytorch#273

manager allreduce expects a tensor but we're passing a list

Reviewed By: amirafzali, d4l3k

Differential Revision: D83353505

fbshipit-source-id: ea5bdee40f9de22720a3135198df478237f1405d

Author: tushar00jain

torchft/process_group.py

@@ -1221,11 +1221,13 @@ def __init__(self, manager: "Manager") -> None:
         self._manager = manager
 
     def allreduce(self, tensors: List[torch.Tensor], opts: object) -> Work:
+        assert len(tensors) == 1
+
         if isinstance(opts, ReduceOp):
-            return self._manager.allreduce(tensors, reduce_op=opts)
+            return self._manager.allreduce(tensors[0], reduce_op=opts)
 
         if isinstance(opts, AllreduceOptions):
-            return self._manager.allreduce(tensors, reduce_op=opts.reduceOp)
+            return self._manager.allreduce(tensors[0], reduce_op=opts.reduceOp)
 
         assert False, "unreachable"