Skip to main content

Module paged

qualia_core_db::inference::runtime::kv

Module paged 

Source
Expand description

Fixed-capacity paged KV planning.

Construction may allocate bounded vectors. Append, fork, copy-on-write, release, and GPU table export never grow them, so the decode and scheduling paths do not touch the global allocator.

Structs§

BlockPool
Fixed-capacity physical-page allocator with reference counts.
GpuBlockTablePlan
Cold-owned immutable block table uploaded by a prepared GPU plan.
PagedKvConfig
SequenceBlockTable
One sequence’s logical-page ownership. Capacity is fixed at construction.

Enums§

AttentionError
CopyOnWrite
PoolError
TableError

Constants§

DEFAULT_BLOCK_SIZE
INVALID_BLOCK

Functions§

fill_identity_block_table
Fill the production single-sequence GPU table with deterministic identity pages.
paged_gqa_attention_into
Scalar oracle for the production paged GQA online-softmax kernel.