ReadonlydeviceThe GPU's architecture identifier as MLX reports it (for example applegpu_g16s).
ReadonlydraftReadonlygrammarGrammar-forced continuations enter the KV without one forward per token (jump-forward decoding, default off).
ReadonlygreedyEvery speculative row drafts by argmax instead of the request sampler's
coupled draw (default off). The draft moves acceptance and draft cost, not
the output distribution (see generation/speculative/draft-policy).
ReadonlykvThe storage and reads of every attention cache the graph's makeCache composes.
ReadonlymaxThe most rows one forward carries (--batch, mlx-lm's --batch-size).
ReadonlymixedOne forward carries a decode group and a prefill group where the graph declares mixed rows (default off).
ReadonlymixedPrompt tokens one mixed forward carries beside its decode rows (default 256).
ReadonlypagedFull-attention keys and values held in pages of blockSize positions
(--paged-kv), or null. Every request's state takes the paged lego in place
of the plain full-attention storage the graph composes; its pages store
what kv names.
ReadonlyprefillThe prefill chunk size, in prompt tokens, the composition is built for. The prefill planner may hand the graph fewer rows than this as a prompt's tail. Its per-prompt halving for a graph with a bounded prefill workspace is the planner's business and does not change what the composition built.
ReadonlyprefillA prompt's last token prefills alone, as the first one-position step (mlx-lm's convention, default on); off, the final chunk prefills whole and token zero samples from it.
ReadonlyvectorA decode step whose live rows are all plain greedy samples them with one argmax over the batch instead of one sampler call per row (default on): the same per-row arithmetic in one graph. Off, every row samples alone.
The facts that fix one served configuration, resolved once by the application before the loader builds the graph and consumed as data. Every field holds a resolved value; nothing here is "unset" or "auto".