DLM Proposer + AR Verifier — runnable KV-cache-saving framework#2

Draft

FluffyAIcode wants to merge 11 commits into

mainfrom

AgentMemory/dlm-proposer-kv-cache-runtime-8e7f

Commits on May 18, 2026

Add runnable DLM-proposer + AR-verifier KV-cache framework

cursoragent
and
FluffyAIcode
committed
Fix NBT accounting: split persistent KV from transient activation

cursoragent
and
FluffyAIcode
committed
Expand 'NBT' acronym to 'Net Bytes per Token' everywhere

cursoragent
and
FluffyAIcode
committed

Commits on May 22, 2026

Add local-inference-engine architecture (no PagedAttention, fixed-slab pool)

cursoragent
and
FluffyAIcode
committed
Review pass on architecture doc + code cleanup

cursoragent
and
FluffyAIcode
committed
Phase 1 of local inference engine: setup scripts + 100% test coverage on kv_cache_proposer/

cursoragent
and
FluffyAIcode
committed
Mac M4 24GB Phase 1 test results

authored and
fluffy314
committed
Fix Phase 1 Mac failure: pre-download HF cache + connectivity probe

cursoragent
and
FluffyAIcode
committed
Fix version-detection bug: mlx exposes no __version__ attribute

cursoragent
and
FluffyAIcode
committed
Phase B: SparseLogitsProposer (lm_head only at masked positions)

cursoragent
and
FluffyAIcode
committed
Mac M4 Phase B bench results

authored and
fluffy314
committed