Deepcopy output node metadata (#95426)

Fixes #ISSUE_NUMBER

Pull Request resolved: pytorch/pytorch#95426
Approved by: https://github.com/SherlockNoMad

torch/fx/graph.py

@@ -796,8 +796,9 @@ def __deepcopy__(self, memo=None) -> 'Graph':
         output_vals = g.graph_copy(self, val_map=memo, return_output_node=True)
         g._codegen = copy.deepcopy(self._codegen)
         assert isinstance(output_vals, tuple)
-        output_val, old_output_val = output_vals
-        g.output(output_val, type_expr=getattr(old_output_val, 'type', None))
+        output_val, old_output_node = output_vals
+        new_output_node = g.output(output_val, type_expr=getattr(old_output_node, 'type', None))
+        new_output_node.meta = copy.copy(old_output_node.meta)
         return g
 
     @compatibility(is_backward_compatible=True)