Hi, sorry for adding to an old post. I recently encountered the same error while playing with TIR scheduling primitives. The discussion in the thread makes sense to me and I understand the fundamental difficulty of having nested parallelization.
One thing that confuses me is that according to the TIR paper (ASPLOS 2023) TIR has a correctness guarantee on schedule primitives which means if a schedule can be applied without error, it’s valid. In this case, nested parallel is not allowed but the error is thrown all the way until codegen, rather than TIR scheduling time. Adding a check here seems to better cope with the TIR correctness claims. Would be happy to send a PR on this. cc @junrushao @yzh119