Handle reduce-window in bfloat16 propagation
If the reduce computation's is precision preserving (e.g. max) then the reduce window is precision preserving as well so we can propagate bfloat16 through it. PiperOrigin-RevId: 220783083
Loading
Please sign in to comment