Vision-only prompt (back-compat wrapper over buildMultimodalPrompt —
behavior byte-identical to the pre-audio builder: the returned imageMask
serves as BOTH the bidirectional-attention mask and the per-layer-input
zeroing mask, which forwardEmbeddings derives from bidir when no
separate multimodal mask is passed).
Vision-only prompt (back-compat wrapper over buildMultimodalPrompt — behavior byte-identical to the pre-audio builder: the returned imageMask serves as BOTH the bidirectional-attention mask and the per-layer-input zeroing mask, which forwardEmbeddings derives from
bidirwhen no separate multimodal mask is passed).