🚀 The feature, motivation and pitch
lift_to_meta (usually required by shape propagation) and canonicalize_graph don't work well together since canonicalize_graph may change the state dict.
Let's reconsider moving shape propagation out of canonicalize_graph and provide a separate utility for it.
The shape propagation is only used here and here anyway (i.e. inside interface.py when the config flag is set.
This should be a fairly easy change. Let's just make sure that we don't cause any regressions in the expected behavior of the utility
Original slack discussion: https://nvidia.slack.com/archives/C08T55LHSG4/p1756141384472509?thread_ts=1755978699.817669&cid=C08T55LHSG4