Function hc_post_batch

server/src/deepseek4/deepseek4_graph.cpp:2109–2130

    const bool use_flash = attention_impl != DeepSeek4AttentionImpl::Explicit &&
                           (n_tokens > 1 || indexer_topk != nullptr);
    if (use_flash) {
        if (exact_two_band) {
            // A larger scheduling band must retain the numerical topology of
            // two 2K requests. Prefix queries use the first band's F32 raw KV;
            // suffix queries see its final SWA tail after the same F16 cache
            // round-trip. HC, projections and MoE still run once over the full
            // token batch, avoiding a second expert-weight sweep.
            const int first_count = DS4_NUMERICAL_PREFILL_BAND;
            const int second_count = n_tokens - first_count;
            const int first_comp = ratio > 0
                ? ds4_comp_rows_used(lc.comp_kv, lc.n_comp, ratio,
                                     kv_start + first_count - 1)
                : 0;
            const int second_comp = n_comp_live;
            const int second_prior_count = std::min(first_count, w.n_swa);

            auto view_kv = [&](int first, int count) {
                return ggml_view_2d(
                    ctx, kv, head_dim, count, kv->nb[1],
                    (size_t) first * kv->nb[1]);
            };
            auto append_comp = [&](ggml_tensor * raw, int count) {
                if (count <= 0 || !comp_kv_source) return raw;
                ggml_tensor * comp = ggml_view_2d(
                    ctx, comp_kv_source, head_dim, count,