Module kernels 

Primitives shared by more than one model generation.

A function lives here only when every caller uses it with identical semantics and it bakes in no geometry. Anything conditional on generation stays with that generation.

Constants

EPS

Epsilon used by both _rms_unit and _zcrms, inside the square root.

SINKHORN_ITERS

Sinkhorn iterations upstream runs.

Functions

rms_unit

decode._rms_unit — RMS normalisation with no learned scale.

rms_unit_to

rms_unit into a separate destination.

sinkhorn

architecture._sinkhorn — alternating log-domain row/column normalisation of an n x n matrix, returning exp(log_K).