Qualcomm AI Engine Direct - Fix Full (pytorch#19359)

### Summary
Based on reported issue:
pytorch#19179
2 mathematically equivalent model in nn.Module lowered both lowered to
QNN, however, 1 accuracy is good while another is bad. Debugged with
qnn_intermediate_debugger and targeted that accuracy drop is possibly
caused by `full` operation. Took a look and noticed the op builder did
not take layout transform into consideration.
<img width="1785" height="1171" alt="image"
src="https://github.com/user-attachments/assets/c8ca4cf6-b242-4171-a8ee-3595e0bcd376"
/>


### Test plan
Passing tests under test_qnn_delegate.py and fix issue in:
pytorch#19179

Author: winskuo-quic

backends/qualcomm/builders/op_full.py

@@ -25,8 +25,9 @@ def define_node(
         node: torch.fx.Node,
         nodes_to_wrappers: Dict[torch.fx.Node, PyQnnManager.TensorWrapper],
     ) -> PyQnnManager.PyQnnOpWrapper:
+        tensor_shape = list(self.get_tensor(node, node).shape)
         out_tensor = torch.full(
-            node.args[0], node.args[1], dtype=node.meta["val"].dtype
+            tensor_shape, node.args[1], dtype=node.meta["val"].dtype
         )
 
         # since we can derive the constant value of current op in AoT stage

backends/qualcomm/builders/op_full_like.py

@@ -25,8 +25,8 @@ def define_node(
         node: torch.fx.Node,
         nodes_to_wrappers: Dict[torch.fx.Node, PyQnnManager.TensorWrapper],
     ) -> PyQnnManager.PyQnnOpWrapper:
-        in_tensor = node.args[0].meta["val"]
-        ref_tensor = torch.zeros(in_tensor.shape, dtype=in_tensor.dtype)
+        in_tensor = self.get_tensor(node, node)
+        ref_tensor = torch.zeros(list(in_tensor.shape), dtype=in_tensor.dtype)
         out_tensor = torch.full_like(ref_tensor, node.args[1])
 
         # since we can derive the constant value of current op in AoT stage

backends/qualcomm/tests/models.py

@@ -875,6 +875,31 @@ def forward(self, x):
         return self.second(self.first(x))
 
 
+class ConvFull(torch.nn.Module):
+    def __init__(self, fill, full_shape):
+        super().__init__()
+        self.conv = torch.nn.Conv2d(8, 16, 3, padding=1)
+        self.fill = fill
+        self.full_shape = full_shape
+
+    def forward(self, x):
+        y = self.conv(x)
+        c = torch.full(self.full_shape, self.fill, dtype=y.dtype)
+        return torch.cat([y, c], dim=1)
+
+
+class ConvFullLike(torch.nn.Module):
+    def __init__(self, fill):
+        super().__init__()
+        self.conv = torch.nn.Conv2d(8, 16, 3, padding=1)
+        self.fill = fill
+
+    def forward(self, x):
+        y = self.conv(x)
+        c = torch.full_like(y, self.fill)
+        return torch.cat([y, c], dim=1)
+
+
 class ConvTranspose1dSingle(torch.nn.Module):
     def __init__(self, bias=True, dilation=1):
         super().__init__()

backends/qualcomm/tests/test_qnn_delegate.py

@@ -2351,6 +2351,17 @@ def test_qnn_backend_einsum_outer_product_relu(self):
         )
         self.lower_module_and_test_output(module, sample_input)
 
+    def test_qnn_backend_full_layout_transformed(self):
+        full_shape = (1, 16, 4, 6)
+        module = ConvFull(0.5, full_shape)  # noqa: F405
+        sample_input = (torch.randn(1, 8, 4, 6),)
+        self.lower_module_and_test_output(module, sample_input)
+
+    def test_qnn_backend_full_like_layout_transformed(self):
+        module = ConvFullLike(0.5)  # noqa: F405
+        sample_input = (torch.randn(1, 8, 4, 6),)
+        self.lower_module_and_test_output(module, sample_input)
+
     # TODO: Create a new UT class for passes specific checks
     def test_qnn_backend_lift_add_tensor(self):
         module = LiftAddTensor()  # noqa: F405
@@ -5270,6 +5281,19 @@ def test_qnn_backend_einsum_outer_product_relu(self):
         module = self.get_qdq_module(module, sample_input)
         self.lower_module_and_test_output(module, sample_input)
 
+    def test_qnn_backend_full_layout_transformed(self):
+        full_shape = (1, 16, 4, 6)
+        module = ConvFull(0.5, full_shape)  # noqa: F405
+        sample_input = (torch.randn(1, 8, 4, 6),)
+        module = self.get_qdq_module(module, sample_input)
+        self.lower_module_and_test_output(module, sample_input)
+
+    def test_qnn_backend_full_like_layout_transformed(self):
+        module = ConvFullLike(0.5)  # noqa: F405
+        sample_input = (torch.randn(1, 8, 4, 6),)
+        module = self.get_qdq_module(module, sample_input)
+        self.lower_module_and_test_output(module, sample_input)
+
     @unittest.skipIf(is_qnn_sdk_version_less_than("2.35"), "UT pass after QNN 2.35")
     def test_qnn_backend_masked_softmax(self):
         if self.enable_x86_64: