The read kernels of a composition's affine-quantized caches. tiled-causal
is OptiQ's rule: a window through the N-tiled fused port while the cache's own
mask is plain causal (where the port takes the configuration), every other
read through the unfused port. unfused is mlx-lm's port for every read.
--fused-sdpa on composes the first and off the second; unset, a per-layer
scheme (--kv-quant config) composes tiled-causal and a uniform one
unfused, which keeps each scheme's oracle parity.
The read kernels of a composition's affine-quantized caches.
tiled-causalis OptiQ's rule: a window through the N-tiled fused port while the cache's own mask is plain causal (where the port takes the configuration), every other read through the unfused port.unfusedis mlx-lm's port for every read.--fused-sdpa oncomposes the first andoffthe second; unset, a per-layer scheme (--kv-quant config) composestiled-causaland a uniform oneunfused, which keeps each scheme's oracle parity.