The recurrent multi-token-prediction block split from a Qwen release, its phases named by the caller like a graph's: the draft step indexed by the rows it drafts for.
Readonly
One position per row (a draft step, the round's bridge or full-accept row), by the rows' row count.
Several positions per row (prompt rows, a verified continuation).
The recurrent multi-token-prediction block split from a Qwen release, its phases named by the caller like a graph's: the draft step indexed by the rows it drafts for.