The checkpoint-native multi-token-prediction row of a graph that declares nativeDraft, over
compressed (MLA) attention state. The graph owns the numerics and the head's weights; the draft
source owns sampling and state retention.
One draft step: ids [B, 1] (one token per sequence) over each row's anchor
hidden [B, 1, hidden], appended to cache; the caller owns the output
[B, 1, hidden].
A window of two or more tokens per sequence, ids [B, L] over hidden
[B, L, hidden], appended to cache; the caller owns the output. The caller
names the phase from the window it built: one token per sequence is decode.
The checkpoint-native multi-token-prediction row of a graph that declares
nativeDraft, over compressed (MLA) attention state. The graph owns the numerics and the head's weights; the draft source owns sampling and state retention.