ut3_concatenate_weights#

t3toolbox.backend.ut3_operations.ut3_concatenate_weights(weights_A, weights_B)#
def ut3_concatenate_weights(
        weights_A: UT3WeightsData,  # (tucker_weight_supercore, tt_weight_supercore, masks)
        weights_B: UT3WeightsData,  # (tucker_weight_supercore, tt_weight_supercore, masks)
) -> UT3WeightsData:                # per-edge concatenation: padded widths ADD, masks concatenate

Per-edge concatenation of two uniform weights – the + / direct-sum combine, where ranks add. Uniform twin of t3_concatenate_weights, and the weight-side partner of ut3_add (which concatenates the object’s rank masks in exactly the same way).

Both the padded supercores and the masks concatenate on the last axis, so the output’s padded width is nA + nB and its real rank is rank_A + rank_B.

The output mask can go gappy, and that is correct (docs/uniform_masks_vs_ranks.md): if A’s mask has slack (real rank below its padded width), the concatenation leaves a hole between A’s real slots and B’s block. The data stays put – no compaction – and ut3weights_to_t3weights gathers through the holes. Re-canonicalize with the SVD if a prefix form is wanted.

Neither operand’s real slots move, so no masking is needed: the supercores are copied, not reduced, and each output slot comes from exactly one input slot (garbage padding stays in the padding).

Parameters:

  • weights_A (UT3WeightsData)

  • weights_B (UT3WeightsData)

Return type:

UT3WeightsData