nx_kv_arena.nx

nx_kv_arena.nx -- PagedAttention page pool (bits-up). vLLM's PagedAttention (UC Berkeley, SOSP'23) is the killer LLM-serving algorithm: instead of one contiguous K/V buffer per sequence, split into fixed-size PAGES. Each sequence owns a list of page indices; pages can be SHARED across sequences for prefix caching (e.g., shared system prompt = one cached page); free pages return to the pool. vLLM published 24x throughput over HuggingFace TGI on the same hardware. THIS primitive is the page-pool substrate. K/V tensor data itself is stored elsewhere (composes with shipped nx_kv_cache for the tensor mechanics). This is the ALLOCATOR + INDEX layer that makes paged-attention possible. V1 scope: - Fixed pool of NX_KV_ARENA_MAX_PAGES pages - Each page is a fixed-size byte block (caller picks page_bytes at arena_new time; typical 4 KiB for 16 tokens of f16 KV per head) - Sequence holds a list of page indices it owns (no sharing in V1; prefix caching V2 adds refcount-on-pages) - Pool exhaustion -> NULL append; caller handles (caller's responsibility to evict via nx_lysosome or yield via nx_yield_protocol) - Canary-bracketed for tamper detection Deferred to V2: - Cross-sequence page sharing (refcount per page; prefix cache) - Page eviction policy (LRU? attention-class? composes with nx_attention_class) - Dynamic page allocation (grow arena on demand) - Multi-tier pages (GPU VRAM vs RAM vs NVMe; composes with conductor cardinal) genealogy_id: vllm_paged_attention_2023_sosp + s_lora_punica_2023 + cardinal_2026-05-20_elder_ai_off_docker + cardinal_2026-05-20_bits_up_nishi_not_linux lineage_id: substrate_kv_arena_v1