Do not constant prop for mutable buffers. (#7779)

Summary:

Before this change, operations on mutable buffers will be constant prop and replaced with a constant tensor. This change avoids constant prop for mutable buffers.

Differential Revision: D68371513

Author: hsharma35

exir/passes/constant_prop_pass.py

@@ -16,7 +16,6 @@
     get_buffer,
     get_lifted_tensor_constant,
     get_param,
-    is_buffer,
     is_lifted_tensor_constant,
     is_param,
 )
@@ -77,6 +76,15 @@ def get_data(
         return const_node_to_tensor[arg]
     return None
 
+def is_constant_buffer(program: "ExportedProgram", node: torch.fx.Node) -> bool:
+    """Checks if the given node is a constant buffer."""
+
+    if node.target not in program.graph_signature.inputs_to_buffers:
+        return False
+    fqn = program.graph_signature.inputs_to_buffers[node.target]
+    # if the buffer is mutated then record that
+    return fqn not in program.graph_signature.buffers_to_mutate.values()
+
 
 def get_constant_placeholder_dict(
     exported_program: ExportedProgram,
@@ -85,15 +93,12 @@ def get_constant_placeholder_dict(
     Returns a dictionary of placeholder node -> constant tensor.
     """
     const_node_to_tensor: OrderedDict[torch.fx.Node, torch.Tensor] = OrderedDict()
-    for node in exported_program.graph.nodes:
-        if node.op != "placeholder":
-            continue
-
+    for node in exported_program.graph.find_nodes(op="placeholder"):
         if is_param(exported_program, node):
             const_node_to_tensor[node] = cast(
                 torch.Tensor, get_param(exported_program, node)
             )
-        elif is_buffer(exported_program, node):
+        elif is_constant_buffer(exported_program, node):
             const_node_to_tensor[node] = cast(
                 torch.Tensor, get_buffer(exported_program, node)
             )

exir/tests/test_passes.py

@@ -1594,6 +1594,34 @@ def forward(self, x):
             gm.code
         )
 
+    def test_constant_prop_pass_for_mutable_buffers(self) -> None:
+        def count_adds(gm: torch.fx.GraphModule) -> int:
+            return len(
+                gm.graph.find_nodes(
+                    op="call_function", target=exir_ops.edge.aten.add.Tensor
+                )
+            )
+
+        class MutableStateModule(torch.nn.Module):
+            def __init__(self):
+                super().__init__()
+                self.register_buffer("state", torch.zeros(1))
+
+            def forward(self, x):
+                x = x + self.state
+                # Add 1 (constant) to state.
+                self.state.add_(1)
+                return x
+
+        edge_manager = to_edge(
+            export(MutableStateModule(), (torch.zeros(1),), strict=True)
+        )
+        self.assertEqual(count_adds(edge_manager.exported_program().graph_module), 2)
+        edge_manager._edge_programs["forward"] = constant_prop_pass(
+            edge_manager._edge_programs["forward"]
+        )
+        self.assertEqual(count_adds(edge_manager.exported_program().graph_module), 2)
+
     def test_constant_prop_pass_for_no_grad(self) -> None:
         class LSTM(torch.nn.Module):
             def __init__(self, input_size, hidden_size, num_layers):