Arm backend: Propagate and validate device in TOSAQuantizer

backends/arm/_passes/decompose_add_sub_alpha_pass.py

@@ -1,4 +1,4 @@
-# Copyright 2025 Arm Limited and/or its affiliates.
+# Copyright 2025-2026 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.
@@ -76,7 +76,11 @@ def call_operator(self, op, args, kwargs, meta, updated: bool | None = False):
         lhs, rhs = args
 
         alpha_full = super().call_operator(
-            full_op, ((1,), float(alpha)), {}, meta, updated=True
+            full_op,
+            ((1,), float(alpha)),
+            {"device": meta["val"].device},
+            meta,
+            updated=True,
         )
         scaled_rhs = super().call_operator(
             mul_op,

backends/arm/_passes/decompose_avg_pool2d_pass.py

@@ -1,4 +1,4 @@
-# Copyright 2025 Arm Limited and/or its affiliates.
+# Copyright 2025-2026 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.
@@ -51,6 +51,7 @@ def call_operator(self, op, args, kwargs, meta):
         full_op, cat_op, avgpool_op, mul_op = get_decomposition(op)
 
         x = args[0]
+        full_kwargs = {"device": x.data.device}
         kernel_h, kernel_w = args[1]
         kernel_size = kernel_h * kernel_w
         if len(args) > 2 and args[2] is not None:
@@ -75,7 +76,7 @@ def call_operator(self, op, args, kwargs, meta):
         if count_include_pad and pad_w > 0:
             pre_pad_shape = [n, c, h, pad_w]
             pre_pad = super().call_operator(
-                full_op, (pre_pad_shape, 0.0), kwargs, meta, updated=True
+                full_op, (pre_pad_shape, 0.0), full_kwargs, meta, updated=True
             )
 
             if ceil_mode and divisor_override is None:
@@ -88,7 +89,7 @@ def call_operator(self, op, args, kwargs, meta):
             if post_pad_w > 0:
                 post_pad_shape = [n, c, h, post_pad_w]
                 post_pad = super().call_operator(
-                    full_op, (post_pad_shape, 0.0), kwargs, meta, updated=True
+                    full_op, (post_pad_shape, 0.0), full_kwargs, meta, updated=True
                 )
                 cat_nodes = [pre_pad, x, post_pad]
             else:
@@ -103,7 +104,7 @@ def call_operator(self, op, args, kwargs, meta):
         if count_include_pad and pad_h > 0:
             pre_pad_shape = [n, c, pad_h, w + pad_w + post_pad_w]
             pre_pad = super().call_operator(
-                full_op, (pre_pad_shape, 0.0), kwargs, meta, updated=True
+                full_op, (pre_pad_shape, 0.0), full_kwargs, meta, updated=True
             )
 
             if ceil_mode and divisor_override is None:
@@ -116,7 +117,7 @@ def call_operator(self, op, args, kwargs, meta):
             if post_pad_h > 0:
                 post_pad_shape = [n, c, post_pad_h, w + pad_w + post_pad_w]
                 post_pad = super().call_operator(
-                    full_op, (post_pad_shape, 0.0), kwargs, meta, updated=True
+                    full_op, (post_pad_shape, 0.0), full_kwargs, meta, updated=True
                 )
                 cat_nodes = [pre_pad, x, post_pad]
             else:
@@ -142,7 +143,7 @@ def call_operator(self, op, args, kwargs, meta):
             override_multiplier = super().call_operator(
                 full_op,
                 ([1, 1, 1, 1], kernel_size / divisor_override),
-                kwargs,
+                full_kwargs,
                 meta,
                 updated=True,
             )

backends/arm/_passes/decompose_div_tensor_mode.py

@@ -1,4 +1,4 @@
-# Copyright 2025 Arm Limited and/or its affiliates.
+# Copyright 2025-2026 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.
@@ -78,7 +78,7 @@ def call_operator(self, op, args, kwargs, meta):
             zero = super().call_operator(
                 opset["full"],
                 args=((1,) * len(meta["val"].size()), 0.0),
-                kwargs={"dtype": torch.float32},
+                kwargs={"dtype": torch.float32, "device": meta["val"].device},
                 meta=meta,
                 updated=True,
             )

backends/arm/_passes/decompose_groupnorm_pass.py

@@ -1,4 +1,4 @@
-# Copyright 2025 Arm Limited and/or its affiliates.
+# Copyright 2025-2026 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.
@@ -92,9 +92,11 @@ def call(self, graph_module: torch.fx.GraphModule):
             if isinstance(meta["val"], tuple):
                 shape = meta["val"][0].size()
                 dtype = meta["val"][0].dtype
+                device = meta["val"][0].device
             else:
                 shape = meta["val"].size()
                 dtype = meta["val"].dtype
+                device = meta["val"].device
             match len(args):
                 # MI profile always provides all the args: x, weight, bias, N, C, HxW, group, eps
                 case 8:
@@ -156,7 +158,7 @@ def call(self, graph_module: torch.fx.GraphModule):
                     graph_module.graph,
                     full_op,
                     args=(epsilon_reshaped_shape, eps),
-                    kwargs={"dtype": dtype},
+                    kwargs={"dtype": dtype, "device": device},
                     from_node=node,
                 )
                 add0 = create_node(

backends/arm/_passes/decompose_layernorm_pass.py

@@ -1,4 +1,4 @@
-# Copyright 2024-2025 Arm Limited and/or its affiliates.
+# Copyright 2024-2026 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.
@@ -108,9 +108,11 @@ def call(self, graph_module: torch.fx.GraphModule):
             if isinstance(meta["val"], tuple):
                 shape = meta["val"][0].size()
                 dtype = meta["val"][0].dtype
+                device = meta["val"][0].device
             else:
                 shape = meta["val"].size()
                 dtype = meta["val"].dtype
+                device = meta["val"].device
             rank = len(shape)
             dims = list(range(-1, -1 * (n_dims + 1), -1))
             dims = [dim % rank for dim in dims]
@@ -142,7 +144,7 @@ def call(self, graph_module: torch.fx.GraphModule):
                     graph_module.graph,
                     full_op,
                     args=(epsilon_reshaped_shape, epsilon),
-                    kwargs={"dtype": dtype},
+                    kwargs={"dtype": dtype, "device": device},
                     from_node=node,
                 )
                 add0 = create_node(

backends/arm/_passes/decompose_leaky_relu_pass.py

@@ -1,4 +1,4 @@
-# Copyright 2025 Arm Limited and/or its affiliates.
+# Copyright 2025-2026 Arm Limited and/or its affiliates.
 # All rights reserved.
 #
 # This source code is licensed under the BSD-style license found in the
@@ -57,6 +57,7 @@ def call_operator(self, op, args, kwargs, meta):
         x = args[0]
         slope = args[1] if len(args) > 1 else 0.01
         dtype = x.node.meta["val"].dtype
+        device = x.node.meta["val"].device
         clamp, full, mul, add = _get_leaky_relu_ops(op)
         op1 = super().call_operator(
             op=clamp, args=(x, 0, None), kwargs=kwargs, meta=meta
@@ -67,7 +68,7 @@ def call_operator(self, op, args, kwargs, meta):
         op3 = super().call_operator(
             op=full,
             args=(x.node.meta["val"].shape, slope),
-            kwargs={"dtype": dtype},
+            kwargs={"dtype": dtype, "device": device},
             meta=meta,
         )
         op4 = super().call_operator(op=mul, args=(op3, op2), kwargs=kwargs, meta=meta)

backends/arm/_passes/decompose_meandim_pass.py

@@ -1,4 +1,4 @@
-# Copyright 2024-2025 Arm Limited and/or its affiliates.
+# Copyright 2024-2026 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.
@@ -167,7 +167,11 @@ def _reduce_by_sum(self, op, input_node, dims, meta, dtype):
 
         sum = super().call_operator(sum_op, (input_node, dims, True), {}, meta, True)
         full = super().call_operator(
-            full_op, ([1] * len(output_shape), 1 / N), {"dtype": dtype}, meta, True
+            full_op,
+            ([1] * len(output_shape), 1 / N),
+            {"dtype": dtype, "device": input_node.data.device},
+            meta,
+            True,
         )
         if (quant_ops := get_quantization(input_node.node.target)) is not None:
             # Insert Q and DQ nodes after full op.

backends/arm/_passes/decompose_var_pass.py

@@ -1,4 +1,4 @@
-# Copyright 2024-2025 Arm Limited and/or its affiliates.
+# Copyright 2024-2026 Arm Limited and/or its affiliates.
 # All rights reserved.
 #
 # This source code is licensed under the BSD-style license found in the
@@ -98,7 +98,7 @@ def call_operator(self, op, args, kwargs, meta):
         full = super().call_operator(
             full_op,
             ([], 1 / max(0, N - correction)),
-            {"dtype": dtype},
+            {"dtype": dtype, "device": x.data.device},
             meta,
             True,
         )

backends/arm/_passes/scalars_to_attribute_pass.py

@@ -1,4 +1,4 @@
-# Copyright 2024-2025 Arm Limited and/or its affiliates.
+# Copyright 2024-2026 Arm Limited and/or its affiliates.
 # All rights reserved.
 #
 # This source code is licensed under the BSD-style license found in the
@@ -49,13 +49,14 @@ def call(self, graph_module: GraphModule) -> PassResult:
                     shape = get_first_fake_tensor(arg).shape
                     biggest_rank = max(biggest_rank, len(shape))
 
+            output_fake_tensor = get_first_fake_tensor(n)
             new_args: list[Node | int] = []
             for arg in n.args:
                 if isinstance(arg, Node):
                     new_args.append(arg)
                     continue
                 if isinstance(arg, int) and not torch.is_floating_point(
-                    get_first_fake_tensor(n)
+                    output_fake_tensor
                 ):
                     new_args.append(arg)
                     continue
@@ -64,7 +65,8 @@ def call(self, graph_module: GraphModule) -> PassResult:
                 get_new_attr_name = get_new_attr_name_with_prefix(prefix)
                 tensor_constant_name = get_new_attr_name(graph_module)
                 float_tensor = torch.tensor(
-                    float(cast(Union[int, float], arg))
+                    float(cast(Union[int, float], arg)),
+                    device=output_fake_tensor.device,
                 ).reshape((1,) * biggest_rank)
                 graph_module.register_buffer(tensor_constant_name, float_tensor)
                 fake_mode = n.meta["val"].fake_mode

backends/arm/quantizer/arm_quantizer.py

@@ -1,6 +1,6 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
-# Copyright 2024-2025 Arm Limited and/or its affiliates.
+# Copyright 2024-2026 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.
@@ -597,8 +597,50 @@ def _annotate_io(
                 mark_node_as_annotated(node)
 
     def validate(self, model: GraphModule) -> None:
-        """TODO: Implement validation of annotated graph for TOSA backend."""
-        pass
+        """Validate the quantization results. Currently, this includes:
+            - Ensure tensor inputs to each operator live on the same device.
+
+        Args:
+            model (GraphModule): GraphModule being validated.
+        Raises:
+            ValueError: If tensor inputs for any operator span more than one
+                device.
+        """
+        for node in model.graph.nodes:
+            if node.op != "call_function":
+                continue
+
+            devices = set()
+            for arg_node in node.all_input_nodes:
+                meta_val = arg_node.meta.get("val", None)
+                if meta_val is None:
+                    continue
+                if isinstance(meta_val, (tuple, list)):
+                    for tensor in meta_val:
+                        devices.add(
+                            str(
+                                getattr(
+                                    tensor,
+                                    "device",
+                                    f"Could not get device from {tensor}",
+                                )
+                            )
+                        )
+                else:
+                    devices.add(
+                        str(
+                            getattr(
+                                meta_val,
+                                "device",
+                                f"Could not get device from {meta_val}",
+                            )
+                        )
+                    )
+
+                if len(devices) > 1:
+                    raise ValueError(
+                        f"Quantizer detected operator {node.name} with different device inputs: {devices}."
+                    )
 
     def quantize_with_submodules(
         self,

backends/arm/test/quantizer/test_tosa_quantizer_validate.py

@@ -0,0 +1,55 @@
+# 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.
+
+import pytest
+import torch
+
+from executorch.backends.arm.quantizer import TOSAQuantizer
+from executorch.backends.arm.tosa import TosaSpecification
+from torch.fx import symbolic_trace
+
+
+def _annotate_placeholders_with_devices(gm, device_map):
+    for node in gm.graph.nodes:
+        if node.op == "placeholder":
+            device = device_map[node.target]
+            node.meta["val"] = torch.empty(1, device=device)
+
+
+def _get_quantizer():
+    return TOSAQuantizer(TosaSpecification.create_from_string("TOSA-1.0+INT"))
+
+
+class TwoIndependentAdds(torch.nn.Module):
+
+    def forward(self, x: torch.Tensor, y: torch.Tensor):
+        return x + 1, y + 1
+
+
+class CrossDeviceAdd(torch.nn.Module):
+
+    def forward(self, x: torch.Tensor, y: torch.Tensor):
+        return x + y
+
+
+def test_validate_allows_different_devices_across_operators():
+    gm = symbolic_trace(TwoIndependentAdds())
+    _annotate_placeholders_with_devices(
+        gm, {"x": torch.device("cpu"), "y": torch.device("meta")}
+    )
+
+    quantizer = _get_quantizer()
+    quantizer.validate(gm)
+
+
+def test_validate_rejects_mixed_devices_within_operator():
+    gm = symbolic_trace(CrossDeviceAdd())
+    _annotate_placeholders_with_devices(
+        gm, {"x": torch.device("cpu"), "y": torch.device("meta")}
+    )
+
+    quantizer = _get_quantizer()
+    with pytest.raises(ValueError, match="Quantizer detected operator"):
+        quantizer.validate(gm)

backends/transforms/replace_scalar_with_tensor.py

@@ -1,6 +1,6 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
-# Copyright 2025 Arm Limited and/or its affiliates.
+# Copyright 2025-2026 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.
@@ -59,7 +59,10 @@ def get_replacement(self, op, args, kwargs, meta):
                         (1,),
                         args[1],
                     ),
-                    kwargs={"dtype": args[0].to_tensor().dtype},
+                    kwargs={
+                        "dtype": args[0].to_tensor().dtype,
+                        "device": args[0].to_tensor().device,
+                    },
                     meta=meta,
                 ),
                 # Other args.