I need to add that we have measured the compilation time for different model sizes. We think that in recent TVM versions it scales in a non-linear manner (quadratically?) with increasing height of model graph.
The code for reference is a naive LSTM approach https://gist.github.com/grwlf/3a8190e2d1ea4bb9bcb021a2b50b7e27 where num_timesteps is the number of LSTM cells to unroll.