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

dequant_mxfp6

def dequant_mxfp6[fmt: FP6Format, *, SF_VECTOR_SIZE: Int = Int(32)](ctx: DeviceContext, output: TileTensor[Storage=output.Storage, address_space=output.address_space, linear_idx_type=output.linear_idx_type], input: TileTensor[Storage=input.Storage, address_space=input.address_space, linear_idx_type=input.linear_idx_type], scales: TileTensor[Storage=scales.Storage, address_space=scales.address_space, linear_idx_type=scales.linear_idx_type], num_rows: Int, num_cols: Int, pdl_level: PDLLevel = PDLLevel())

Dequantizes MXFP6 packed weights to BF16 or FP8.

Parameters:

• ​fmt (FP6Format): The FP6 encoding of the packed input (E2M3 or E3M2).
• ​SF_VECTOR_SIZE (Int): Elements each E8M0 block scale covers (defaults to 32).

Args:

• ​ctx (DeviceContext): Device context for kernel launch.
• ​output (TileTensor[Storage=output.Storage, address_space=output.address_space, linear_idx_type=output.linear_idx_type]): Output tensor [num_rows, num_cols] of bfloat16 or float8_e4m3fn.
• ​input (TileTensor[Storage=input.Storage, address_space=input.address_space, linear_idx_type=input.linear_idx_type]): Input tensor [num_rows, num_cols * 6 // 8] of uint8, holding four packed FP6 codes per three bytes.
• ​scales (TileTensor[Storage=scales.Storage, address_space=scales.address_space, linear_idx_type=scales.linear_idx_type]): Scale tensor [num_rows, num_cols // SF_VECTOR_SIZE] of float8_e8m0fnu.
• ​num_rows (Int): Number of rows (N dimension for weights).
• ​num_cols (Int): Number of columns (K dimension, unpacked element count).
• ​pdl_level (PDLLevel): PDL optimization level for kernel launch.