Chatting with @eellison, there are some invariants that would be nice to have:

• targets are only HOP or OpOverload. targets cannot be arbitrary python functions or OpOverloadPacket.
• OpOverloads's args are the non-kwonly-args and the kwargs are the kw-only args, or some other standard convention. This is true after tracing but might not be true after graph passes!
• node.meta["val"] are all populated. Or maybe we should make FakeTensorUpdater faster (https://github.com/pytorch/pytorch/blob/main/torch/_inductor/fx_utils.py#L55), it is currently O(# of nodes in graph)
• Inductor pattern matcher's replacement graph is assumed to be functional, but we don't check this. Also related: Improving inductor pattern matcher's replacement graph assumptions #134363

cc @ezyang @chauhang @penguinwu @voznesenskym @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @bdhirsh