Composition gate for token fast-forwarding, enforced inside the engine
The serving planner owns request-level eligibility, including explicit
seeds in serial execution and mounted draft models. Shared strict fill
has its own grouped method. The composition's KV scheme is no gate: every
cache lego takes known tokens, one position at a time or through the
graph's committed spans.
grammar: forced-token production is the grammar's job; two producers in
one iteration is forbidden (asserted in the loop).
logprobs/top_logprobs: injected tokens are never sampled, so they have
no distribution row (same rule as shouldUseGrammarJump).
Composition gate for token fast-forwarding, enforced inside the engine The serving planner owns request-level eligibility, including explicit seeds in serial execution and mounted draft models. Shared strict fill has its own grouped method. The composition's KV scheme is no gate: every cache lego takes known tokens, one position at a time or through the graph's committed spans.