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

convert_ref_scales_to_mxfp8_format

def convert_ref_scales_to_mxfp8_format[MType: CoordLike, NType: CoordLike, KType: CoordLike, ref_a_scales_tt_layout: TensorLayout, ref_b_scales_tt_layout: TensorLayout, a_scales_tt_layout: TensorLayout, b_scales_tt_layout: TensorLayout, //, ref_scales_type: DType, scales_type: DType, *, REF_BLOCK_SIZE: Int, SF_VECTOR_SIZE: Int](m: MType, n: NType, k: KType, ref_a_scales: TileTensor[ref_scales_type, ref_a_scales_tt_layout, Engine=ref_a_scales.Engine, linear_idx_type=ref_a_scales.linear_idx_type], ref_b_scales: TileTensor[ref_scales_type, ref_b_scales_tt_layout, Engine=ref_b_scales.Engine, linear_idx_type=ref_b_scales.linear_idx_type], a_scales: TileTensor[scales_type, a_scales_tt_layout, Engine=a_scales.Engine, linear_idx_type=a_scales.linear_idx_type], b_scales: TileTensor[scales_type, b_scales_tt_layout, Engine=b_scales.Engine, linear_idx_type=b_scales.linear_idx_type])

Converts reference float32 block scales into the 5D MXFP8 E8M0 scale-factor layout.

Reads the per-block float32 reference scales for the A (M x K) and B (N x K) operands, converts each to float8_e8m0fnu, and writes them into the corresponding 5D scale-factor tensors expected by block-scaled matmul kernels.

Parameters:

• ​MType (CoordLike): CoordLike type carrying the M dimension size.
• ​NType (CoordLike): CoordLike type carrying the N dimension size.
• ​KType (CoordLike): CoordLike type carrying the K dimension size.
• ​ref_a_scales_tt_layout (TensorLayout): TensorLayout of the 2D reference A scales.
• ​ref_b_scales_tt_layout (TensorLayout): TensorLayout of the 2D reference B scales.
• ​a_scales_tt_layout (TensorLayout): TensorLayout of the 5D output A scales.
• ​b_scales_tt_layout (TensorLayout): TensorLayout of the 5D output B scales.
• ​ref_scales_type (DType): Element type of the reference scales (must be float32).
• ​scales_type (DType): Element type of the output scales (must be float8_e8m0fnu).
• ​REF_BLOCK_SIZE (Int): Block size (in elements) used by the reference scales.
• ​SF_VECTOR_SIZE (Int): Number of elements each scale factor covers in the output layout.

Args:

• ​m (MType): M dimension size of the operands.
• ​n (NType): N dimension size of the operands.
• ​k (KType): K dimension size of the operands.
• ​ref_a_scales (TileTensor[ref_scales_type, ref_a_scales_tt_layout, Engine=ref_a_scales.Engine, linear_idx_type=ref_a_scales.linear_idx_type]): 2D float32 reference scales for the A operand, indexed as [k // REF_BLOCK_SIZE, m].
• ​ref_b_scales (TileTensor[ref_scales_type, ref_b_scales_tt_layout, Engine=ref_b_scales.Engine, linear_idx_type=ref_b_scales.linear_idx_type]): 2D float32 reference scales for the B operand, indexed as [n // REF_BLOCK_SIZE, k // REF_BLOCK_SIZE].
• ​a_scales (TileTensor[scales_type, a_scales_tt_layout, Engine=a_scales.Engine, linear_idx_type=a_scales.linear_idx_type]): Mutable 5D output tensor receiving the converted A scales.
• ​b_scales (TileTensor[scales_type, b_scales_tt_layout, Engine=b_scales.Engine, linear_idx_type=b_scales.linear_idx_type]): Mutable 5D output tensor receiving the converted B scales.