Function build_output_context

gguf-tools/deepseek4-quantize.c:1577–1604

            fprintf(stderr, "generate_expert_tensor: %s %d %s %d/%d experts\n",
                    j->expert.scope == EXP_SCOPE_MTP ? "stage" : "layer",
                    j->expert.layer, j->wid, done, j->n_experts);
        }
        pthread_mutex_unlock(&j->lock);
    }
    return NULL;
}

static byte_buf generate_expert(st_db *db, const char *gguf_name, const tensor_meta *tmpl,
                                ds4q_type target, int n_experts, int n_threads,
                                const imatrix_store *imatrix) {
    expert_tensor e = parse_expert_tensor(gguf_name);
    if (!e.is_expert) die("not an expert tensor");
    if (target != DS4Q_TYPE_MXFP4 && !is_quantizable_target(target)) {
        die("unsupported expert target type");
    }
    const char *wid = expert_part_name(e.part);
    const int64_t ncols = tmpl->ne[0];
    const int64_t nrows = tmpl->ne[1];
    const size_t per_expert = (size_t)nrows * ds4q_row_size(target, ncols);
    byte_buf out = { .size = per_expert * (size_t)n_experts, .data = xmalloc(per_expert * (size_t)n_experts) };
    ds4q_quantize_init(target);
    int worker_count = n_threads > 0 ? n_threads : 8;
    if (worker_count < 1) worker_count = 1;
    if (worker_count > n_experts) worker_count = n_experts;
    fprintf(stderr, "generate_expert_tensor: %s %d %s using %d worker%s\n",
            e.scope == EXP_SCOPE_MTP ? "stage" : "layer",
            e.layer, wid, worker_count, worker_count == 1 ? "" : "s");
    expert_job job = {
        .db = db, .gguf_name = gguf_name, .tmpl = tmpl, .target = target,
        .n_experts = n_experts, .imatrix = imatrix, .expert = e, .wid = wid,
        .ncols = ncols, .nrows = nrows, .per_expert = per_expert, .out = &out,