nx_f32_kv_cache.nx

nx_f32_kv_cache.nx -- bits-up f32 KV cache substrate. L7 / L8 composition brick. Closes the autoregressive-decode O(n^2) → O(n) gap: instead of recomputing K/V from the prompt every step, cache them per layer and only project the new token. Cache layout (linear i64 arrays of f32 bits): cache_K: [n_layers, max_seq_len, kv_dim] row-major cache_V: [n_layers, max_seq_len, kv_dim] kv_dim = n_kv_heads * head_dim seq_len: scalar count of filled rows (advanced once per fwd pass) Per-layer append + advance pattern: For each layer L: 1. Project x -> K_new, V_new (caller does this) 2. nx_f32_kv_cache_append_layer(cache, L, K_new, V_new, n_new) copies K_new/V_new into cache_K[L, seq_len:seq_len+n_new, :] 3. Read K_all = nx_f32_kv_cache_get_K(cache, L) Read V_all = nx_f32_kv_cache_get_V(cache, L) Both are full pointers (caller uses seq_len + n_new rows) 4. Attention over those (n_new + cached) rows Once ALL layers done: nx_f32_kv_cache_advance(cache, n_new) Increments seq_len by n_new. Substrate-honest: this v1 stores f32 outputs (post-projection). Production KV caches sometimes use int8 or even Q4_K -- queued as v2 quantized-KV variant once a quant primitive lands. genealogy_id: standard_kv_cache_canon + decoder_lm_decode_pattern lineage_id: substrate_f32_kv_cache_v1