attentionMask property
The effective attention mask over this result's seq axis, when the
graph knows one (design D-T3's silent-regression fix). A runTokens
that pads a fixed-shape graph MUST pad ids/mask/type-ids together and
echo the padded mask here, so the masked mean-pool excludes exactly the
padding it added — never the tokenizer's mask, which describes the
UNPADDED input. One over a dynamic-shape graph (no padding) may pass the
input mask through unchanged or omit it — finalizeEmbedding's caller
then falls back to the tokenizer's mask.
Implementation
final List<int>? attentionMask;