Bias the choice of experts while retaining original affinities

Read the equation

An expert-specific correction vector is added to the sigmoid affinities for ranking. This addition can change which groups and experts are selected. The later mixture weights are gathered from the original affinities, so the correction bias does not directly enter their numerical values.

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

2.  scores = sigmoid(logits)                                        -> (T, 256)
      Sigmoid, not softmax. Scores do not sum to 1 and are not a distribution.

  3.  scores_for_choice = scores + e_score_correction_bias            -> (T, 256)
      The F32 (256,) bias, broadcast over the token dimension.

  4.  group_scores = scores_for_choice.view(T, 8, 32).topk(2, -1)[0].sum(-1)
                                                                      -> (T, 8)

scores_for_choice = scores + e_score_correction_bias            -> (T, 256)

Related builds