[QNN] Fix LiftConstantScalarOperands to handle aten.pow.Scalar (pytorch#18994)

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Author: KevinUW114514

backends/qualcomm/_passes/lift_constant_scalar_operands.py

@@ -48,6 +48,7 @@ class TensorOpInfo:
     aten.sub.Scalar: TensorOpInfo(aten.sub.Tensor, False, False),
     aten.sub.Tensor: TensorOpInfo(aten.sub.Tensor, False, False),
     aten.pow.Tensor_Scalar: TensorOpInfo(aten.pow.Tensor_Tensor, False, False),
+    aten.pow.Scalar: TensorOpInfo(aten.pow.Tensor_Tensor, False, False),
     # The scalar number arg[1] is missing when using default. Result in a corner case to deal
     aten.leaky_relu.default: TensorOpInfo(aten.prelu.default, True, False),
     aten.leaky_relu_.default: TensorOpInfo(aten.prelu.default, True, False),
@@ -86,11 +87,13 @@ def _build_tensor_constant(
     ) -> TensorConstant:
         # For dtype, in some cases, we cannot use node.args[0] as scalar dtype.
         # Ex: Where op args[0] can be bool, however, we probably want args[1] and args[2] to be dtype same as node.meta["val"] instead of bool type
+        first_arg = node.args[0]
         tensor = torch.tensor(
             const_val,
             dtype=(
-                node.args[0].meta["val"].dtype
-                if not is_float_tensor(node)
+                first_arg.meta["val"].dtype
+                if isinstance(first_arg, fx.Node)
+                and not is_float_tensor(node)
                 and (info := SCALAR_OPS.get(node.target))
                 and not info.use_self_dtype
                 else node.meta["val"].dtype

backends/qualcomm/tests/models.py

@@ -1829,6 +1829,15 @@ def forward(self, x):
         return torch.pow(x, self.exponent)
 
 
+class PowScalar(torch.nn.Module):
+    def __init__(self, base=2.0):
+        super().__init__()
+        self._base = base
+
+    def forward(self, x):
+        return torch.ops.aten.pow.Scalar(self._base, x)
+
+
 class PReLUDefault(torch.nn.Module):
     def __init__(self):
         super().__init__()

backends/qualcomm/tests/test_qnn_delegate.py

@@ -1713,6 +1713,26 @@ def test_qnn_backend_pow_tensor_scalar(self):
                         index += 1
                         self.lower_module_and_test_output(module, sample_input)
 
+    def test_qnn_backend_pow_scalar(self):
+        test_comb = [
+            {
+                QCOM_MODULE: [
+                    PowScalar(),  # base=2.0, default  # noqa: F405
+                    PowScalar(3.0),  # base=3.0, common case  # noqa: F405
+                    PowScalar(9),  # base=9, integer exp case  # noqa: F405
+                    PowScalar(0.5),  # base=0.5, fractional case  # noqa: F405
+                ],
+                QCOM_SAMPLE_INPUTS: [(torch.rand(10, 10) + 0.1,)],
+            },
+        ]
+        index = 0
+        for comb in test_comb:
+            for module in comb[QCOM_MODULE]:
+                for sample_input in comb[QCOM_SAMPLE_INPUTS]:
+                    with self.subTest(i=index):
+                        index += 1
+                        self.lower_module_and_test_output(module, sample_input)
+
     def test_qnn_backend_prelu(self):
         test_comb = [
             {
@@ -4229,6 +4249,27 @@ def test_qnn_backend_pow_tensor_scalar(self):
                         qdq_module = self.get_qdq_module(module, sample_input)
                         self.lower_module_and_test_output(qdq_module, sample_input)
 
+    def test_qnn_backend_pow_scalar(self):
+        test_comb = [
+            {
+                QCOM_MODULE: [
+                    PowScalar(),  # base=2.0, default  # noqa: F405
+                    PowScalar(3.0),  # base=3.0, common case  # noqa: F405
+                    PowScalar(9),  # base=9, integer exp case  # noqa: F405
+                    PowScalar(0.5),  # base=0.5, fractional case  # noqa: F405
+                ],
+                QCOM_SAMPLE_INPUTS: [(torch.rand(10, 10) + 0.1,)],
+            },
+        ]
+        index = 0
+        for comb in test_comb:
+            for module in comb[QCOM_MODULE]:
+                for sample_input in comb[QCOM_SAMPLE_INPUTS]:
+                    with self.subTest(i=index):
+                        index += 1
+                        qdq_module = self.get_qdq_module(module, sample_input)
+                        self.lower_module_and_test_output(qdq_module, sample_input)
+
     def test_qnn_backend_prelu(self):
         test_comb = [
             {