Expand description
2026-09-25: The paged KV block pool, host-side KV dequantization, the KV spill-file manager and the radix-tree prefix cache.
Owner: cache. Invariants: none beyond the types.
Modulesยง
- kv_
cache - 2026-09-25: The paged KV cache: the
KvCacheDtypestorage formats, the per-layer block geometry inKvCacheConfig, and thePagedKvCacheblock pool with its per-block reference counts. - kv_
dequant - 2026-09-25: Host-side dequantization of paged KV cache blocks to BF16.
- kv_
spill - 2026-09-25: Swap files for the
--swap-space-gbspill of sequence state. - radix_
tree - 2026-09-25: The radix-tree prefix cache (
PrefixCacheimplementation): KV block reuse plus the SSM snapshot index.