nx_f32_attn_multi.nx

nx_f32_attn_multi.nx -- bits-up f32 multi-token attention kernel. L7 / L8 composition brick. The prerequisite for both prefill (processing a whole prompt) and KV cache decode (attending to growing cache). Composes nx_f32_dot + nx_f32_mul + nx_f32_softmax + nx_f32_add. No libm. Algorithm: Inputs: Q [n_tokens_q, head_dim] query tensor K [n_tokens_k, head_dim] key tensor (may be larger when attending to a KV cache) V [n_tokens_k, head_dim] value tensor For each query token i in 0..n_tokens_q: For each key token j in 0..n_tokens_k: scores[j] = dot(Q[i,:], K[j,:]) * attn_scale (if causal: positions j > q_pos_i are masked with -inf) probs = softmax(scores) For each output dim d: out[i, d] = sum_j probs[j] * V[j, d] For causal LM during prefill: q_pos_i = i (query i can attend to keys 0..=i). For causal LM during KV-cache decode: q_pos_i = n_tokens_k - n_tokens_q + i (query at end of cache, attends to all prior). K^T transpose is handled by index arithmetic (we access K row-by- row in its natural [n_tokens_k, head_dim] layout); no physical transpose needed. genealogy_id: vaswani_2017_attention + causal_mask_standard lineage_id: substrate_f32_attn_multi_v1