Fix ChannelsLastTaggedReshapePass crash on ops with list-typed args (#18958)

The `input_to_nhwc` method crashes with `AttributeError:
'immutable_list' object has no attribute 'graph'` when lowering
dynamically quantized models containing ops like `cat` whose first
argument is a list of tensors.

Two fixes:
1. Add `isinstance(args[0], torch.fx.Node)` guard in the dynamic input
trace-back loop to stop when args[0] is not a Node.
2. Handle list-typed `node.args[0]` at the call site by iterating over
each element and converting them individually to NHWC.
Fixes #18944

Author: kirklandsign

backends/xnnpack/_passes/channels_last_tagged_reshape_pass.py

@@ -398,8 +398,11 @@ def input_to_nhwc(
             is_dynamic_input = is_dynamic_qdq(input_node)
 
             if is_dynamic_input:
-                # Trace back to original source node
-                while getattr(input_node, "args", None):
+                # Trace back to original source node. Stop if args[0] is not
+                # a Node (e.g., immutable_list from cat).
+                while getattr(input_node, "args", None) and isinstance(
+                    input_node.args[0], torch.fx.Node
+                ):
                     input_node = input_node.args[0]
 
             with graph_module.graph.inserting_after(input_node):
@@ -505,7 +508,13 @@ def call(self, graph_module: torch.fx.GraphModule):  # noqa: C901
             elif self.requires_nhwc_input(node):
                 # Nodes which enter this branch are ones that require their
                 # first input to be nhwc. This makes this node's output nhwc too
-                self.input_to_nhwc(graph_module, node.args[0], node)
+                if isinstance(node.args[0], (list, tuple)):
+                    # Ops like cat have a list of tensors as args[0].
+                    for arg in node.args[0]:
+                        if isinstance(arg, torch.fx.Node):
+                            self.input_to_nhwc(graph_module, arg, node)
+                else:
+                    self.input_to_nhwc(graph_module, node.args[0], node)
                 for input_node in node.all_input_nodes[1:]:
                     if (
                         input_node.op == "placeholder"

backends/xnnpack/test/passes/test_channels_last_tagged_reshape.py

@@ -633,3 +633,35 @@ def forward(self, x):
         ]
         self.assertEqual(1, len(view_nodes))
         self.assertTrue(ChannelsLastTaggedReshapePass(None).is_nchw_node(view_nodes[0]))
+
+    class ConvCat(torch.nn.Module):
+        def __init__(self):
+            super().__init__()
+            self.conv1 = torch.nn.Conv2d(3, 16, 3, padding=1)
+            self.conv2 = torch.nn.Conv2d(3, 16, 3, padding=1)
+
+        def forward(self, x):
+            return torch.cat([self.conv1(x), self.conv2(x)], dim=1)
+
+    def test_fp32_conv_cat_immutable_list(self):
+        model = self.ConvCat().eval()
+        x = torch.randn(1, 3, 8, 8)
+        self.run_tester(model, (x,))
+
+    def test_dq_conv_cat_immutable_list(self):
+        model = self.ConvCat().eval()
+        x = torch.randn(1, 3, 8, 8)
+        (
+            Tester(model, (x,))
+            .quantize(
+                Quantize(
+                    quantization_config=get_symmetric_quantization_config(
+                        is_dynamic=True
+                    )
+                )
+            )
+            .export()
+            .to_edge()
+            .run_passes(self.PassStage)
+            .run_method_and_compare_outputs()
+        )