Understand export and lowering

The central export and lowering sequence is:

    

        
        
example_inputs = (torch.randn(*spec["input"]["shape"]),)

with torch.inference_mode():
    exported = export(model, example_inputs)

edge = to_edge_transform_and_lower(
    exported,
    partitioner=[XnnpackPartitioner()],
)
program = edge.to_executorch()

    

This is the same export-and-lower sequence used for GPT-2 and DistilBERT in Module 1. The difference lies in the graph. Transformer graphs contain attention, masks, reshapes, indexing, and normalization around their main compute. MobileNetV2 is dominated by convolution patterns that can form a more compact XNNPACK region.