nx_f32_mha_multi.nx

nx_f32_mha_multi.nx -- bits-up f32 multi-head + multi-token attention. L7 / L8 composition brick. Combines the head-decomposition pattern from nx_f32_mha.nx (multi-head + GQA) with the multi-token causal kernel from nx_f32_attn_multi.nx into the full attention forward. This is the actual attention kernel for: - Prefill (n_tokens = prompt length, causal mask) - KV-cache decode (n_tokens = 1 new + cache; future Task #13) - Non-causal encoder attention (causal = 0) Algorithm: Q_all = matmul(x, W_q) [n_tokens, n_heads*head_dim] K_all = matmul(x, W_k) [n_tokens, n_kv_heads*head_dim] V_all = matmul(x, W_v) [n_tokens, n_kv_heads*head_dim] For each query head h in 0..n_heads: kv_head = h / group_size Slice Q_h, K_h, V_h from the column range [h*head_dim, (h+1)*head_dim] for queries (or kv_head*head_dim for K,V). attn_h = nx_f32_attn_multi(Q_h, K_h, V_h, n_tokens, n_tokens, head_dim, causal, attn_scale) Scatter attn_h back into attn_out_all at columns [h*head_dim, (h+1)*head_dim]. out = matmul(attn_out_all, W_o) [n_tokens, hidden_dim] Slicing is via copy in v1. v2 will use stride-aware kernels to avoid the copy (perf lift queued). genealogy_id: vaswani_2017_mha + ainslie_2023_gqa + causal_mask_canon lineage_id: substrate_f32_mha_multi_v1