vLLM KV Cache Layout: block ID → byte address

vLLM KV cache layout: block ID → byte address

One flat byte array; block table maps logical → physical b, no addresses. Each same-shape layer group gets its own KVCacheTensor view:
addr = offset + l × layer_stride + b × block_stride + slot × row_width

layout

block_size

num_blocks

scenarioDeepSeek sparse attention: MLA (compressed latent KV) 656 B/token × 2 layers + Indexer (sparse index keys) 132 B/token × 2 layers, sharing one block table

KVCacheTensor params per spec

spec	offset	layer_stride	block_stride	row width	page size

Memory map (flat byte array, to scale)

Hover for byte range; green = selected token.

MLA L0 MLA L1 Indexer L0 Indexer L1

Look up a token's address

token t

layer