dcaballe wrote: Perhaps we can also add a callback function so that the user can decide when an op should be linearized or not. That should give us the level of flexibility that we need. https://github.com/llvm/llvm-project/pull/81159