From 213c547ec3d740cc2b91dda29200aa9513c190ed Mon Sep 17 00:00:00 2001
From: Jacob Szwejbka <jakeszwe@meta.com>
Date: Sat, 22 Aug 2026 15:14:37 -0700
Subject: [PATCH] Ignore symbolic metadata for lifted constants (#22056)

Summary:

D116957343 began lifting `get_attr` tensor constants after edge transforms. Constant-backed `FakeTensor` values can retain symbolic-looking shape metadata even though pass replay uses their attached concrete constants, causing a false symbolic-input validation failure.

Exclude constant-backed `FakeTensor` values from symbolic runtime-input snapshots while retaining validation for actual symbolic inputs.

Reviewed By: shoumikhin

Differential Revision: D117081177
---
 exir/pass_base.py             |  2 ++
 exir/tests/test_pass_infra.py | 22 ++++++++++++++++++++++
 2 files changed, 24 insertions(+)

diff --git a/exir/pass_base.py b/exir/pass_base.py
index 7d026acf878..6071aae2be8 100644
--- a/exir/pass_base.py
+++ b/exir/pass_base.py
@@ -121,6 +121,8 @@ def _leaf_symbolic_snapshot(value: Argument) -> Any:
         return scalar_snapshot
 
     if isinstance(value, FakeTensor):
+        if value.constant is not None:
+            return None
         dims = []
         has_symbolic_dim = False
         for dim in value.shape:
diff --git a/exir/tests/test_pass_infra.py b/exir/tests/test_pass_infra.py
index 59406b13f8f..16ed5af4180 100644
--- a/exir/tests/test_pass_infra.py
+++ b/exir/tests/test_pass_infra.py
@@ -24,6 +24,7 @@
 from executorch.exir.passes import ScalarToTensorPass
 from executorch.exir.passes.pass_registry import PassRegistry
 from executorch.exir.program import to_edge
+from torch._subclasses.fake_tensor import FakeTensor
 from torch.export import Dim, export, ExportedProgram
 from torch.export.graph_signature import InputKind, InputSpec, TensorArgument
 from torch.fx.passes.infra.pass_base import PassBase, PassResult
@@ -513,6 +514,27 @@ def placeholder(
             any(dim is not None for dim in self._symbolic_input_shape(new_input))
         )
 
+    def test_export_pass_ignores_symbolic_metadata_for_constant_input(self) -> None:
+        graph_module = self._export_dynamic_graph_module()
+        original_input = self._find_input_node(graph_module)
+        original_value = original_input.meta["val"]
+        self.assertIsInstance(original_value, FakeTensor)
+        assert isinstance(original_value, FakeTensor)
+        constant = torch.randn(2, 3)
+        original_value.constant = constant
+
+        new_graph_module = ExportPass()(graph_module).graph_module
+        new_input = self._find_input_node(new_graph_module)
+        new_value = new_input.meta["val"]
+
+        self.assertIsInstance(new_value, FakeTensor)
+        assert isinstance(new_value, FakeTensor)
+        self.assertIs(new_value.constant, constant)
+        self.assertEqual(
+            self._symbolic_input_shape(new_input),
+            self._symbolic_input_shape(original_input),
+        )
+
     def test_export_pass_rejects_collapsed_symbolic_input_metadata(self) -> None:
         class CollapseSymbolicInputPass(ExportPass):
             def placeholder(