Tuple is not variadic once it’s created, but I feel for autocast_promote we are not going to update the number of arguments/outputs once we inserted the node, so that should be fine?
Having said that, I actually hate having a Tuple there. It makes profiling/aliasing analysis messy in the unfortunate PyTorch world. Is there any benefit in having a tuple as output comparing to your prototype above? I think it’s quite common in the current JIT IR to have variadic operations (guard / requires_grad_check e.t.c.)