FakeTensorUpdater expects graphs where all callable nodes can be called with FakeTensor. It appears that there is a class of functions with the attribute _inductor_lowering_function set that explicitly expect IRNode inputs, but can appear in graphs passed to FakeTensorUpdater. This should be looked into, since we can't update these nodes appropriately (possibly leading to silent incorrectness). See #159523 (comment) for additional context.

cc @ezyang @gchanan @zou3519 @kadeng @msaroufim @chauhang @penguinwu @eellison @bdhirsh