invert_mxfp8_scale#

nvmath.linalg.advanced.helpers.matmul.invert_mxfp8_scale(mx_scales)[source]#

This function is experimental and potentially subject to future changes.

Compute a reciprocal of MXFP8 block scale.

Parameters:

mx_scales – MXFP8 block scale tensor (UE8M0 format).

Returns:

Tensor of the same shape as mx_scales with exponents replaced by the reciprocals.