Sampler options for the DRAFT of a request. greedy is the composition's
greedyDraft: argmax drafts, which drop the key coupling.
MLX_BUN_SPEC_DRAFT_TEMPERATURE=<t> samples the draft at its own
temperature: a head flatter than its target matches more often when
sharpened (t < request temperature); t=0 is argmax.
MLX_BUN_SPEC_DRAFT_FILTER=0 drops top-p/top-k/min-p/XTC from the draft:
the coupled inverse-CDF draw over the full vocabulary, a quarter of the
per-step sampling cost, which at low temperature rarely leaves the
filtered set.
Sampler options for the DRAFT of a request.
greedyis the composition'sgreedyDraft: argmax drafts, which drop the key coupling.MLX_BUN_SPEC_DRAFT_TEMPERATURE=<t>samples the draft at its own temperature: a head flatter than its target matches more often when sharpened (t < request temperature); t=0 is argmax.MLX_BUN_SPEC_DRAFT_FILTER=0drops top-p/top-k/min-p/XTC from the draft: the coupled inverse-CDF draw over the full vocabulary, a quarter of the per-step sampling cost, which at low temperature rarely leaves the filtered set.