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_mxfp6_even_scale

def compute_mxfp6_even_scale[fmt: FP6Format](max_val: Float32) -> Float8_e8m0fnu

Computes the OCP MXFP6 E8M0 scale using even-mode rounding.

The FP6 generalization of compute_mxfp4_even_scale: it rounds the block maximum at the target's mantissa boundary before deriving the exponent, so a block whose maximum is just under a power-of-two boundary does not waste a whole binade of range on the other 31 elements.

Parameters:

• ​fmt (FP6Format): The FP6 encoding the block will be quantized to.

Args:

• ​max_val (Float32): The largest magnitude in the 32-element block.

Returns:

Float8_e8m0fnu: The E8M0 scale for the block.