ReadonlycfgReadonlydraftThe draft block per row count or range of its rows (anchors times columns), indexed by the scheduler: every 4-bit
projection on the kernel projectContext runs for the class; BF16 projections run one matmul at every
class.
ReadonlygammaDraft tokens per block (block size minus the anchor).
ReadonlyprojectThe context projection (fc over the target's taps, then the hidden norm) per row count or range of the rows
projected (B × L: a prompt's rows at prefill, a commit's accepted rows), indexed by the caller
(B × L). A 4-bit fc runs the kernel the target declares for the class or, by the drafter's
own rule, the simdgroup matrix unit (Affine4MmaLinear, where the weight's shape fits it) at 5..8
and MLX's quantized matmul (QuantizedLinear) at every other class; a BF16 fc runs one matmul at every
class. No entry runs before bindTarget.
ReadonlyprojectEach layer's k and v over the context rows (k norm, RoPE), per row count or range of those rows (B × L), indexed
like projectContext and on the kernel it names for the class.
ReadonlytapApply the target's declarations before any projection: its residual basis (null: the trained basis) and the kernels its composition names for the drafter's 4-bit projections (null: the drafter's own rule). The first call applies them; every later call must pass the same declarations.
The columns of a block drafting depth tokens: its anchor and up to depth mask positions, at most the
trained block size.
The DFlash 2 block drafter: target taps → projected context K/V; one greedy block per anchor. It loads in the basis it was trained on and learns the target's when a draft group binds.