Skip to main content

ClikaRT::nn::PagedParams

struct

Header: ClikaRT/nn/kv_cache.h

Shared block-pool geometry for the cache's paged rows. block_size must be a power of two in [16, 512] (a kernel addressing contract); num_blocks is the pool capacity; max_blocks_per_seq caps one sequence's block-table row.

num_blocks budgets the FULL-attention paged rows only. A WINDOWED paged row (KVLayerSpec::window > 0) provisions its own pool on top: max_seqs × (⌈window/block_size⌉ + 1) blocks per distinct window value, a small, deterministic footprint (it can never exhaust), logged at construction so the cache's memory is never under-reported.

Data members​

block_size​

int32_t block_size = 0

tokens per block

Declared in ClikaRT/nn/kv_cache.h, line 160

num_blocks​

int64_t num_blocks = 0

full rows' shared pool capacity in blocks

Declared in ClikaRT/nn/kv_cache.h, line 161

max_blocks_per_seq​

int32_t max_blocks_per_seq = 0

per-sequence block-table ceiling (full rows)

Declared in ClikaRT/nn/kv_cache.h, line 162