@@ -0,0 +1,79 @@
+# 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.
+"""Regression tests for BF16 aten.pad fallback without the TOSA bf16
+extension.
+"""
+
+import pytest
+import torch
+import torch.nn.functional as F
+from executorch.backends.arm.tosa.compile_spec import TosaCompileSpec
+from executorch.backends.arm.tosa.partitioner import TOSAPartitioner
+from executorch.backends.arm.vgf import VgfCompileSpec, VgfPartitioner
+from executorch.exir import EdgeCompileConfig, to_edge_transform_and_lower
+
+
+class Pad(torch.nn.Module):
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return F.pad(x, (1, 1)) * 2
+
+
+def _pad_node(dtype: torch.dtype):
+    ep = torch.export.export(Pad(), (torch.ones(1, 4, 8, dtype=dtype),))
+    node = next(
+        node for node in ep.graph.nodes if node.target == torch.ops.aten.pad.default
+    )
+    return ep, node
+
+
+@pytest.mark.parametrize("dtype", (torch.float32, torch.float16, torch.bfloat16))
+@pytest.mark.parametrize("enable_bf16", (False, True))
+def test_vgf_pad_preservation_respects_bf16_extension(dtype, enable_bf16):
+    spec = "TOSA-1.0+FP+INT+int4+int16"
+    if enable_bf16:
+        spec += "+bf16"
+
+    ep, node = _pad_node(dtype)
+    keep_ops, filter_fn = VgfPartitioner(VgfCompileSpec(spec)).ops_to_not_decompose(ep)
+
+    assert torch.ops.aten.pad.default in keep_ops
+    assert filter_fn is not None
+    # FP32/FP16 are supported by the FP profile. BF16 requires +bf16.
+    assert filter_fn(node) is (dtype != torch.bfloat16 or enable_bf16)
+
+
+@pytest.mark.parametrize("enable_bf16", (False, True))
+def test_tosa_pad_preservation_respects_bf16_extension(enable_bf16):
+    spec = "TOSA-1.0+FP" + ("+bf16" if enable_bf16 else "")
+    ep, node = _pad_node(torch.bfloat16)
+    keep_ops, filter_fn = TOSAPartitioner(TosaCompileSpec(spec)).ops_to_not_decompose(
+        ep
+    )
+
+    assert torch.ops.aten.pad.default in keep_ops
+    assert filter_fn is not None
+    assert filter_fn(node) is enable_bf16
+
+
+def test_vgf_bf16_pad_without_extension_exports_to_pte():
+    # The entire BF16 graph stays portable with
+    # this explicitly BF16-disabled spec.
+    x = torch.randn(1, 4, 8, dtype=torch.bfloat16)
+    ep = torch.export.export(Pad(), (x,))
+    edge = to_edge_transform_and_lower(
+        ep,
+        partitioner=[VgfPartitioner(VgfCompileSpec("TOSA-1.0+FP+INT+int4+int16"))],
+        compile_config=EdgeCompileConfig(_check_ir_validity=False),
+    )
+    targets = [
+        str(node.target)
+        for node in edge.exported_program().graph.nodes
+        if node.op == "call_function"
+    ]
+    assert any("constant_pad_nd" in target for target in targets)
+    assert not any("aten.pad" in target for target in targets)
+    assert not any("executorch_call_delegate" in target for target in targets)
+    # The unfixed version raises: Missing out variants: {'aten::pad'}.
+    edge.to_executorch()
@@ -1009,6 +1009,15 @@ def filter_fn(node: torch.fx.Node) -> bool:
             """
             if _is_custom_partition_op(self._custom_partition_ops, node.target):
                 return True
+            # BF16 pad must decompose when the selected TOSA spec lacks +bf16.
+            # Otherwise it is left on the CPU as aten.pad, which has no
+            # portable out variant. Decomposition produces constant_pad_nd.
+            if (
+                node.target == torch.ops.aten.pad.default
+                and get_first_fake_tensor(node).dtype == torch.bfloat16
+                and not self.tosa_spec.support_extension("bf16")
+            ):
+                return False
             if (
                 node.target in ops_to_not_decompose
                 and get_first_fake_tensor(node).dtype == torch.float64