nx_f32_llama_v4b.nx

nx_f32_llama_v4b.nx -- BATCHED MULTI-SEQUENCE decode twins: one forward step advances M INDEPENDENT paged sequences by one token each (the continuous-batching primitive, vLLM-class serving). WHY IT PAYS: QKV / FFN / lm_head rows are sequence-INDEPENDENT, so M forks batch through the same matmuls -- every weight byte is read once for M tokens instead of once per token (the same amortization that makes chunked prefill fast). Attention is per-sequence: each row r attends its own seqs[r] via nx_f32_attn_with_paged with n_q=1 (100%% reuse). Per-cell matmul math is independent of m, so each row's logits are BIT-IDENTICAL to the m=1 sequential path (gated by nx_batched_gate). Faithful derivative of nx_f32_llama_v4p with EXACTLY these deltas: * pseq -> seqs: *i64 (M pointers to NxPagedSeq), one NEW token per seq * RoPE position per row = seqs[r].seq_len (each row is ITS seq's next) * attention: per-row loop over nx_f32_attn_with_paged(n_q=1) * stack: ensure_append(seq_r, 1) each BEFORE layers; advance each AFTER genealogy_id: kwon_2023_pagedattention + yu_2022_orca_continuous_batching lineage_id: substrate_f32_llama_v4b_v1