Normalize the selected routing affinities

Read the equation

Each gathered affinity is divided by the sum across the selected experts plus 10⁻²⁰. Ignoring that stabilizer, the selected weights sum to one. With the stabilizer retained, the exact sum is S/(S+10⁻²⁰), so it is approximately one for an ordinary positive sum. This is normalization of sigmoid scores, not a softmax of the logits.

Terms & definitions

How it works

Read the definition from its right-hand expression to the quantity on the left. Its surrounding source records the intended input, state, or parameter context. Where the source leaves an operation unspecified, this page preserves that gap.

Assumptions & limits

Source recovery establishes that the expression appears in this record. Its mathematical validity, implementation behavior, and interpretation must be assessed separately. See the linked source and research notes for the checks available for this version.

Original context

SELECTION in steps 3-6 used scores_for_choice, the POST-bias value.
          Selection and weighting use different quantities, deliberately.

  8.  topk_weight = topk_weight / (topk_weight.sum(-1, keepdim=True) + 1e-20)
      topk_weight = topk_weight * 2.5                   # routed_scaling_factor
      Renormalize to sum 1, THEN scale by 2.5.

  Then in DeepseekV3MoE.forward (lines 522-532):

topk_weight = topk_weight / (topk_weight.sum(-1, keepdim=True) + 1e-20)

Related builds