For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Mojo function

compute_mxfp8_block_scale

def compute_mxfp8_block_scale[scales_dtype: DType](group_max: Float32) -> Tuple[Scalar[scales_dtype], Float32, Bool]

Derives an MXFP8 E8M0 block scale and the reciprocal the data path applies.

Parameters:

• ​scales_dtype (DType): E8M0 scale type to emit (float8_e8m0fnu).

Args:

• ​group_max (Float32): Largest absolute value in the 32-element scale block.

Returns:

Tuple[Scalar[scales_dtype], Float32, Bool]: (e8m0_scale, data_multiplier, block_is_dead). Multiply the block's data by data_multiplier and store e8m0_scale. When block_is_dead the caller MUST also zero the data: the block has no E4M3 representation, and a non-finite lane would otherwise store inf * 0.0 = NaN.