Fix output spec + insert clone for constant_prop_pass. (#11209)

Summary:
Pull Request resolved: #11209

In case where a constant propagated node is returned by the exported program, `_validate()` fails with `SpecViolationError` with signature:
```
User output <SOME_CONSTANT_PROP_NODE> is not in the correct order or is not found in the exported program's user_output list.
```
This diff does two things:
1. Update output spec when propagated constants are output of the program.
2. Insert clone op on the constant prop tensor before sending them to output node. This avoid memory planning related errors.

Differential Revision: D75473310

Author: hsharma35

exir/passes/constant_prop_pass.py

@@ -295,6 +295,37 @@ def create_constant_nodes_and_return_specs(
     return name_to_spec_dict
 
 
+def _update_output_node_and_specs(exported_program: ExportedProgram) -> None:
+    """
+    Update the output node and output specs in the exported program.
+    In case a constant node is used as output, we replace it with a clone of the constant node.
+    """
+    # Dict [node.name -> InputSpec]
+    updated_constant_placeholders = get_constant_placeholder_dict(exported_program)
+    output= exported_program.graph.find_nodes(op="output")[0]
+    output_nodes = cast(list[torch.fx.Node], list(output.args[0]))
+    output_specs = exported_program.graph_signature.output_specs
+    assert len(output_nodes) == len(output_specs)
+
+    for i in range(len(output_specs)):
+        out_node = output_nodes[i]
+        if out_node not in updated_constant_placeholders:
+            continue
+
+        with exported_program.graph.inserting_after(out_node):
+            new_node = exported_program.graph.call_function(
+                exir_ops.edge.aten.clone.default, (out_node,)
+            )
+        assert "val" in out_node.meta
+        new_node.meta["val"] = out_node.meta["val"]
+        output_nodes[i] = new_node
+
+        # Update the constant-propagated output node.
+        output_specs[i].arg = TensorArgument(name=output_nodes[i].name)
+
+    output.args = (output_nodes,)
+
+
 def constant_prop_pass(
     exported_program: ExportedProgram,
     custom_skip_targets: Optional[set[EdgeOpOverload]] = None,
@@ -341,12 +372,12 @@ def constant_prop_pass(
 
     # Generate new input spec.
     new_input_specs = []
-    for node in exported_program.graph.nodes:
-        if node.op != "placeholder":
-            continue
+    for node in exported_program.graph.find_nodes(op="placeholder"):
         new_input_specs.append(name_to_spec_dict[node.name])
     exported_program.graph_signature.input_specs = new_input_specs
 
+    _update_output_node_and_specs(exported_program)
+
     # Cleanup the graph.
     exported_program.graph.eliminate_dead_code()
     exported_program.graph_module.recompile()

exir/tests/test_passes.py

@@ -1025,6 +1025,34 @@ def forward(self, x):
         FileCheck().check(
             "executorch_exir_dialects_edge__ops_aten_slice_copy_Tensor"
         ).run(gm.code)
+        
+    def test_constant_prop_for_output(self) -> None:
+        class Add(torch.nn.Module):
+            def forward(self) -> torch.Tensor:
+                return torch.add(torch.tensor(3), torch.tensor(5))
+
+        add = Add()
+
+        edge = to_edge(
+            export(add, (), strict=True),
+            compile_config=EdgeCompileConfig(_skip_dim_order=False),
+        )
+        # Check there is a lifted tensor followed by a to_copy node
+        FileCheck().check("c_lifted_tensor_0").check(
+            "c_lifted_tensor_1"
+        ).run(edge.exported_program().graph_module.code)
+
+        edge._edge_programs["forward"] = constant_prop_pass(edge.exported_program("forward"))
+
+        # Check (c_lifted_tensor_*) nodes are all replaced by _prop_tensor_constant.
+        FileCheck().check_not("c_lifted_tensor_").check(
+            "_prop_tensor_constant"
+        ).run(
+            edge.exported_program().graph_module.code
+        )
+        # Validate that the program successfully passes validation to executorch:
+        edge.exported_program()._validate()
+        edge.to_executorch()
 
     def test_constant_prop_pass_for_add(self) -> None:
         class Add(torch.nn.Module):