Arm backend: Stop removing kwargs in quantizer (#16931)

torchao now allows non-tensor kwargs when quantizing. Currently, I am
not aware of any tensor kwargs so it is probably better to crash and
handle it if we run into them instead of silently not quantizing them.

Add a test to make sure kwargs actually do survive quantization.

Author: Erik-Lundell

backends/arm/quantizer/quantization_annotator.py

@@ -840,16 +840,3 @@ def annotate_graph(  # type: ignore[return]
             _annotate_output(node, quant_properties.quant_output)
 
         mark_node_as_annotated(node)  # type: ignore[attr-defined]
-
-        # Quantization does not allow kwargs for some reason.
-        # Remove from ops we know have and where we know it does not break anything.
-        if node.target in [
-            torch.ops.aten.full_like.default,
-            torch.ops.aten.full.default,
-            torch.ops.aten.full,
-            torch.ops.aten.fill_.Scalar,
-            torch.ops.aten.scalar_tensor.default,
-            torch.ops.aten.zeros.default,
-            torch.ops.aten.ones.default,
-        ]:
-            node.kwargs = {}

backends/arm/test/quantizer/test_preserve_kwargs.py

@@ -0,0 +1,53 @@
+# Copyright 2026 Arm Limited and/or its affiliates.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import Tuple
+
+import pytest
+import torch
+from executorch.backends.arm.test import common
+from executorch.backends.arm.test.tester.test_pipeline import TosaPipelineINT
+from executorch.backends.test.harness.stages.stage import StageType
+
+input_t1 = Tuple[torch.Tensor, int]
+
+exir_op = "executorch_exir_dialects_edge__ops_aten_full_default"
+
+
+class FullLike(torch.nn.Module):
+    """Since full_like is replaced with full, we only need to test on reference model, not FVP."""
+
+    test_parameters = {
+        "full_like_int_val": lambda: (torch.randn(2, 2, 2, 2) * 50, 3),
+        "full_like_float_val": lambda: (torch.randn(2, 4, 5, 2) * 50, 3.2),
+    }
+
+    def forward(self, input_tensor: torch.Tensor, value):
+        # Our backend can't handle tensors without users, which input_tensor doesn't have
+        # when the full_like is converted to a full. Therefore involve it in the output.
+        return input_tensor + torch.full_like(
+            input_tensor, value, dtype=torch.float32, memory_format=torch.channels_last
+        )
+
+
+@common.parametrize("test_data", FullLike.test_parameters)
+def test_preserves_kwargs_tosa_INT(test_data):
+    pipeline = TosaPipelineINT[input_t1](
+        FullLike(),
+        test_data(),
+        aten_op=[],
+        exir_op=exir_op,
+    )
+    pipeline.run()
+
+    # Test that kwarg memory_format survived quantization.
+    graph_module = pipeline.tester.get_artifact(StageType.EXPORT).graph_module
+    nodes = graph_module.graph.nodes
+    for n in nodes:
+        if n.target == torch.ops.aten.full_like.default:
+            assert n.meta["val"].dim_order() == (0, 2, 3, 1)
+            break
+    else:
+        pytest.fail("Did not find torch.ops.aten.full_like")