nx_f32_attn_cached.nx

nx_f32_attn_cached.nx -- bits-up f32 multi-head attention with KV cache. L7 / L8 integration brick. Closes the substrate-side gap between the math primitives (RMSNorm + matmul + softmax + ...) and the real LLM inference pattern (prefill once, then decode token-by- token attending to the growing cache). Caller responsibility: 1. Project x -> Q, K_new, V_new via matmul(x, W_q/k/v). 2. Optionally apply RoPE to each head slice (per token position cache_seq_before + i) via nx_f32_rope_apply_vector. 3. Call this kernel: appends K_new/V_new to cache, attends over full history. 4. Project the concatenated attention output via matmul(., W_o) to get attn_proj. 5. Residual: x = x + attn_proj. Algorithm: 1. Append K_new, V_new to cache at layer_idx. (Caller has NOT yet advanced cache.seq_len.) 2. Total key/value count visible to this layer: total_k = cache.seq_len + n_tokens (the new rows are now in cache at positions seq_len..seq_len+n_tokens-1 ready to be read). 3. For each query head h in 0..n_heads: kv_head = h / group_size Slice Q_h from Q at columns [h*head_dim, (h+1)*head_dim] Read K_all = cache.K[layer_idx] (first total_k rows valid) Slice K_h from K_all (per-token slicing over kv_head's stripe) Same for V_h. attn_h = nx_f32_attn_multi(Q_h, K_h, V_h, n_tokens, total_k, head_dim, causal, attn_scale) Scatter attn_h to attn_concat[:, h*head_dim:(h+1)*head_dim]. 4. The caller advances cache.seq_len AFTER all layers done for this forward pass via nx_f32_kv_cache_advance. Returns: NX_F32_AC_OK on success, error verdicts otherwise. genealogy_id: standard_kv_cache_decode_pattern + mha + gqa lineage_id: substrate_f32_attn_cached_v1