yzhang93 wrote: @fabianmcg could you take another look at this PR? I think there might be other corner cases that need to be addressed. But this at least fixed the issues for convolutions. https://github.com/llvm/llvm-project/pull/149576