Optional ReadonlyadaptiveThe adaptive draft gate's settings per draft provider kind, measured on the hardware it serves (an adaptive drafter of a listed kind gates with them). Absent, or a kind it does not list: the gate's defaults.
Optional ReadonlyadaptiveThe widest adaptive round per draft provider kind, measured on the hardware it serves: an adaptive drafter of a listed kind drafts up to it instead of its own width, unless the request names a width. Absent, or a kind it does not list: the drafter's own width.
Optional ReadonlyappendOptional ReadonlycacheOne cache per cache layer: the caches the composition's KV scheme names
(composedFullAttentionCache, composedSlidingAttentionCache).
Optional ReadonlycompositionThe served configuration the graph was built for; absent for a graph that composes no KV scheme (its caches are its own).
ReadonlyconfigOptional ReadonlydecodeOptional ReadonlyembedUnscaled token embeddings, for media prompts that splice embeddings.
Optional ReadonlyembeddingPrepared embeddings without explicit positions, indexed by total rows.
Optional ReadonlyexpertReadonlygraphOptional ReadonlyknownCommitted known-token spans, declared when the composition qualifies them
(appendKnown then reads committed).
Optional ReadonlylogitsOptional ReadonlyloraThe mounted-adapter selection; absent when the graph mounts none.
Optional ReadonlymixedMixed rows in one forward, declared by a composition that serves them.
Optional ReadonlymixedMixed groups that retain their own projection geometry. The scheduler routes each group independently when a verify group is present.
Optional ReadonlymultiOptional ReadonlymultiOptional ReadonlypromptOptional ReadonlypromptOptional chunk bindings when a file distinguishes chunks from prompt tails.
Optional ReadonlysequenceFull-sequence scoring operations, indexed by total token rows.
Optional ReadonlysequenceOptional ReadonlytrainableWhat the trainer consumes: the head, segmented backward, prefix sharing, gradient checkpointing, flash-attention constraint and denoising objective this graph implements. Undeclared parts make the trainer refuse.
Optional ReadonlyverifyOptional ReadonlyverifyMeasured 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.
ReadonlyweightsOptional ReadonlywiredThe graph's composition measured that its weights are wired for a generation whatever their share of the recommended working set; absent: the share decides.
Optionalbindmedia embeddings/embeddings+positions: bind one request's prepared
embeddings into an input whose prefill is indivisible.
OptionalcompiledcompiledDecode: the graph's compiled decode step, created on first use
and owned by the graph until releaseCompiledDecode.
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.
OptionaldraftPorts over this graph's live caches that draft sources may consume.
OptionalembedOptionalexpertSnapshot of the streamed expert residency, for diagnostics; null when the graph streams none.
Optionalmediamedia: the lazy encoders this graph loads from the checkpoint at modelDir.
OptionalnativenativeDraft: the checkpoint's own multi-token-prediction head, refused when its tier is not loaded.
Optionalpixelmedia pixels: the pixel input a denoising method consumes; null when the checkpoint has none.
OptionalprefillThe vision prefill phase over a prompt's prepared embeddings (borrowed), for
direct generation: image tokens attending each other both ways where
imageMask marks them, ids feeding per-layer inputs; width is the actual number
of the prompt's rows.
OptionalreleaseRetire the compiled step before the graph unloads: its recorded graph borrows the graph's weights and constants. Safe to call when none exists.
What the shared token executors (the batch group, its preparation cohorts, the autoregressive binding) consume from a resident graph, whatever family it is: caches, the named phases (
tokenPhasesOf), weights for the wired-limit scope, and the ports its declaration promises. A graph outside the model registry satisfies it structurally.