Mojo function

compute_dynamic_fp8_scale

compute_dynamic_fp8_scale[out_dtype: DType](row_max: Scalar[dtype], scale_ub: Scalar[dtype]) -> Tuple[Scalar[dtype], Scalar[dtype]]

Compute dynamic FP8 scale factor and its reciprocal from a row max.

Computes scale_factor = min(row_max, scale_ub) / fp8_max and its reciprocal. Does not use math.recip to avoid a reciprocal approximation that gives up too much precision.

Parameters:

out_dtype (DType): The FP8 output dtype (float8_e4m3fn or float8_e4m3fnuz).

Args:

row_max (Scalar): Maximum absolute value across the row/group.
scale_ub (Scalar): Upper bound to clamp the scale factor.

Returns:

Tuple: A tuple of (scale_factor, scale_factor_recip).