Why is the Graph implemented based on networkx.DiGraph instead of FxGraph from torch.dynamo?

Hi, I am curious about the implementation choice of the Graph in NanoFlow. Why does NanoFlow choose to implement a custom Graph based on networkx.DiGraph instead of using the FxGraph from torch.dynamo which can be acquired from the original torch model script (like the one in vllm) through symbolic tracing? What is the underlying considerations?