How does the MLIR that Mojo generates differ from https://openxla.github.io/iree/

Or does Mojo sit on top of iree?

Will the inference engine support multi GPU setups?