Currently for the punet model, we are doing scale math on all layers in f32 because some are sensitive to precision. We should do an experiment to see if there is an overall benefit to switching to lower precision or a different way of doing this. If so, we can spend some time on the simulator attempting to make it work.
Currently for the punet model, we are doing scale math on all layers in f32 because some are sensitive to precision. We should do an experiment to see if there is an overall benefit to switching to lower precision or a different way of doing this. If so, we can spend some time on the simulator attempting to make it work.