ChannelsLastTaggedReshapePass crashes with immutable_list when using dynamic quantization

[jgibson2]

Bug

ChannelsLastTaggedReshapePass.input_to_nhwc crashes with AttributeError: 'immutable_list' object has no attribute 'graph' when lowering a dynamically quantized model to XNNPACK without per_op_mode.

Root Cause

Two issues in channels_last_tagged_reshape_pass.py:

1. input_to_nhwc while loop (line ~402)

if is_dynamic_input:
    while getattr(input_node, "args", None):
        input_node = input_node.args[0]

When tracing back through dynamic quantization nodes, args[0] can be an immutable_list (e.g., from a cat op where the first argument is a list of tensors). The loop assigns this list to input_node, then graph.inserting_after(input_node) crashes because immutable_list has no .graph attribute.

Fix: Add an isinstance check:

if is_dynamic_input:
    while getattr(input_node, "args", None) and isinstance(input_node.args[0], torch.fx.Node):
        input_node = input_node.args[0]

2. call method (line ~508)

self.input_to_nhwc(graph_module, node.args[0], node)

For ops like cat, node.args[0] is an immutable_list of tensors, not a single Node. This is passed directly as input_node to input_to_nhwc.

Fix: Guard with isinstance:

if isinstance(node.args[0], torch.fx.Node):
    self.input_to_nhwc(graph_module, node.args[0], node)

Reproduction

1. Export a model containing torch.cat operations
2. Apply dynamic quantization via XNNPACKQuantizer with get_symmetric_quantization_config(is_per_channel=True, is_dynamic=True)
3. Lower with to_edge_transform_and_lower using XnnpackPartitioner() (without per_op_mode)

The crash occurs during the ChannelsLastTaggedReshapePass preprocessing step. Using per_op_mode=True works around the issue.

Stack Trace

File ".../channels_last_tagged_reshape_pass.py", line 508, in call
    self.input_to_nhwc(graph_module, node.args[0], node)
File ".../channels_last_tagged_reshape_pass.py", line 405, in input_to_nhwc
    with graph_module.graph.inserting_after(input_node):
File ".../torch/fx/graph.py", line 1583, in inserting_after
    if n.graph != self:
AttributeError: 'immutable_list' object has no attribute 'graph'

Environment

• ExecuTorch 1.1.0
• PyTorch 2.9.1
• macOS (Apple Silicon)

Workaround

Use XnnpackPartitioner(per_op_mode=True) to bypass the fusion pass.