Seems like mostly changes for production users vs. researchers. Not that there's any particular feature I'm dying for as a researcher (though a deterministic reduce_sum would be nice). Is pytorch more directly pursuing the "fast iteration, few assumptions" that ML researchers want?
