Single cache
One cache type, one view (offset = 0). Block table maps logical → physical block b, nothing more.addr = offset + l × layer_stride + b × block_stride + slot × row_width
layout
LBNHC (layer-outer)
BLHNC (block-outer)
layers L
2 3
block_size
16 64
num_blocks
3 4 8
scenario MLA (compressed latent KV), row width 656 B/token
KVCacheTensor params
spec offset layer_stride block_stride row width page size
Memory map (flat byte array, to scale)
Hover for byte range; green = selected token.
Look up a token's address
When all layers share one shape, one block table + one KVCacheTensor view (offset = 0) is all you need.
Multiple cache types sharing one block table — see below.