Wouldn’t that be only a problem the other way around, going from float16 -> bfloat16?
Since float16 supports more digits after the decimal points, going bfloat16 -> float16 there shouldn’t be no rounding?

→ View original post on X — @rasbt