Expand description
KV-cache configuration types: quantized cache types and the host-RAM prompt cache setting.
Kept as a standalone, low-complexity module (not folded into
server_config) so cache-related config resolution has one home.
Enums§
- Cache
RamSetting - How to determine the host-RAM prompt cache budget (
--cache-ram). - KvCache
Type - A llama.cpp KV-cache element type, as accepted by
--cache-type-k/--cache-type-v.
Constants§
- DEFAULT_
CACHE_ TYPE_ K - Default K cache type: quantized to roughly halve the KV footprint versus
f16, doubling how much conversation history the RAM/disk prompt caches can hold. - DEFAULT_
CACHE_ TYPE_ V - Default V cache type. Same rationale as
DEFAULT_CACHE_TYPE_K; see the module docs on the Flash Attention requirement for quantized V.