Function sum_gpu_index

llama.cpp:2757–2779

static int64_t sum_gpu_index(struct ggml_tensor * gpu_index) {
    ggml_context * ctx_aux = ggml_init({
        /* mem_size */ 1 << 10,
    });

    GGML_ASSERT(ctx_aux);

    ggml_cgraph * gf = ggml_new_graph_custom(ctx_aux, 1, false);
    ggml_tensor * sum = ggml_sum(ctx_aux, gpu_index);

    ggml_set_name(sum, "gpu_index_sum");
    ggml_build_forward_expand(gf, sum);

    // TODO: +1 worker for GPU under hybrid inference but no use
    // ggml_graph_compute_helper(work_buffer, gf, 2);
    ggml_graph_compute_with_ctx(ctx_aux, gf, 2);

    int32_t sum_val = ggml_get_i32_1d(sum, 0);

    ggml_free(ctx_aux);

    return sum_val;
}

struct llama_gpu_split_loader {
    int n_tensors = 0;