OptionalconfidenceOptionalembedOptionalentropyOptionaleosOptionalfullOptionalmaxOptionalmaxOptionalminOptionalsamplerOptionalseedOptionalstabilityOptionaltemperatureOptionalthresholdOptionaltOptionaltOptionalvisionImage-text-to-text: when set, promptIds are the SPLICED ids (with the
<|image|> run) and the encoder prefills with the merged vision features +
bidirectional overlay instead of the plain text path. Channel-first
[1,3,H,W]. Caller keeps ownership.
Dequantized embedding table for self-conditioning. Borrowed: the caller keeps ownership. Absent → the run dequantizes and owns its own copy.