ReadonlyappendReadonlycacheReadonlyconfigReadonlydecodeReadonlyembedUnscaled token embeddings, for media prompts that splice embeddings.
ReadonlyembeddingPrepared embeddings without explicit positions, indexed by total rows.
ReadonlyfaReadonlyfinalReadonlygraphReadonlygraphReadonlyknownCommitted known-token spans, declared when the composition qualifies them
(appendKnown then reads committed).
ReadonlylmReadonlylogitsReadonlyloraReadonlymediaReadonlymediaReadonlymediaReadonlymixedMixed groups that retain their own projection geometry. The scheduler routes each group independently when a verify group is present.
ReadonlymultiReadonlymultiReadonlyprefixReadonlypromptReadonlypromptOptional chunk bindings when a file distinguishes chunks from prompt tails.
ReadonlysequenceFull-sequence scoring operations, indexed by total token rows.
ReadonlyverifyReadonlyverifyMeasured cost of a verify forward by its row count, relative to one single-row decode step (rows 1 = 1). Declared only by a graph measured on the hardware it serves; speculative depth scheduling reads it.
Readonlyweightsmedia embeddings/embeddings+positions: bind one request's prepared
embeddings into an input whose prefill is indivisible.
Release what the graph built at construction (computed constants, Neural Engine programs, its own weight source where it opened one) once every execution and compiled runner has stopped. A graph that built nothing releases nothing. Its owner (the host, a draft provider) calls it before releasing the weights it supplied.
Ports over this graph's live caches that draft sources may consume.
media: the lazy encoders this graph loads from the checkpoint at modelDir.
A written arrangement qualified on M4 Pro. The loader recommends it; explicit selection remains the caller's choice on any machine.