ReadonlydecodeThe decode read: one query per row.
ReadonlymaskedThe window read once the cache's own state makes its mask an array: padded or unequal rows in a batched cache, the sliding window bound in a rotating cache. The cache picks it on the branch that picks that mask, from state it owns, never from anything the caller passes.
Optional ReadonlymultiThe kernels a row layout reads with unless it holds exactly one row at
its own positions (several rows, or one behind left padding); absent,
these kernels at every row count. Only a set whose kernels take one
sequence has it (the M4 Pro composition's KV4 set, Kv4Head256Cache). The
layout picks it from its own row layout, as it picks maskedWindow from
its own mask, never from the queries it is handed.
ReadonlywindowThe window read (L causal queries per row) while the cache's own mask
is plain causal.
The kernel an affine cache attends with for each read. The cache takes it when it is built and never chooses again:
unfusedAffineKernelsfor a uniform--kv-quant N,tiledCausalAffineKernelsfor--kv-quant config,tiledAffineKernelsfor the opt-in always-tiled composition, or a device-specific cache's own.