Function hc_pre_batch

server/src/deepseek4/deepseek4_graph.cpp:2059–2093

                const int pos_i = kv_start + i;
                float * col = mvals.data() + (size_t) i * n_attn;
                for (int r = 0; r < n_raw; ++r) {
                    const int pos_r = end <= w.n_swa
                        ? r
                        : (end - 1) - ((end - 1 - r) % w.n_swa);
                    if (pos_r > pos_i) col[r] = -1e30f;
                }
                if (n_comp_attn > 0) {
                    const int visible = ds4_comp_rows_used(
                        lc.comp_kv, lc.n_comp, ratio, pos_i);
                    for (int c = visible; c < n_comp_attn; ++c) {
                        col[n_raw + c] = -1e30f;
                    }
                }
                int old_row = 0;
                for (int t = 0; t < n_tokens; ++t) {
                    if (kv_start + t < w.n_swa) continue;
                    if (t <= i) {
                        col[n_raw + n_comp_attn + old_row] = -1e30f;
                    }
                    ++old_row;
                }
            }
            f32_array_inputs->push_back({cmask, std::move(mvals)});
            score_mask = ggml_reshape_2d(ctx, cmask, n_attn, n_tokens);
        }
    }
    const bool direct_indexer_topk = indexer_topk &&
        ds4_env_flag("DFLASH_DS4_DIRECT_INDEXER_TOPK");
    if (indexer_topk) {
        if (!score_mask) {
            score_mask = ggml_new_tensor_2d(
                ctx, GGML_TYPE_F32, n_attn, n_tokens);
            ggml_set_input(score_mask);
            f32_array_inputs->push_back({
                score_mask,
                std::vector<float>((size_t) n_attn * n_tokens, 0.0f),
            });
        }