Function llama_build_graph

llama.cpp:6408–6753

static llm_offload_trie k_offload_func_trie(k_offload_map);

static struct ggml_cgraph * llama_build_graph(
         llama_context & lctx,
     const llama_batch & batch) {
    const auto & model = lctx.model;

    // check if we should build the worst-case graph (for memory measurement)
    const bool worst_case = ggml_allocr_is_measure(lctx.alloc);

    // keep track of the input that has already been allocated
    bool alloc_inp_tokens   = false;
    bool alloc_inp_embd     = false;
    bool alloc_inp_pos      = false;
    bool alloc_inp_KQ_scale = false;
    bool alloc_inp_KQ_mask  = false;
    bool alloc_inp_K_shift  = false;

#ifdef GGML_USE_CUBLAS
    const bool do_offload = true;
#else
    const bool do_offload = true; // TODO: set to false after finishing refactoring
#endif

    int n_non_view = 0; // number of non-view tensors that have been processed by the callback

    // this callback allows us to apply custom logic to each tensor (e.g. ggml-alloc, offloading, etc.)
    // TODO: will be removed with backend v2

    // For sparse deriv, we offload layers from the starting layer to the end.
    // For dense deriv, we offload layers from the end to the starting layer.
    bool offload_starting_layers = lctx.model.sparse_deriv;

    llm_build_cb cb = [&](struct ggml_tensor * cur, const char * name, int il) {
        if (il >= 0) {
            ggml_format_name(cur, "%s-%d", name, il);
        } else {
            ggml_set_name(cur, name);
        }

        //
        // allocate input tensors and set input data
        //
        // TODO: will be removed with backend v2

        if (!alloc_inp_tokens && strcmp(name, "inp_tokens") == 0) {
            ggml_allocr_alloc(lctx.alloc, cur);

            if (!ggml_allocr_is_measure(lctx.alloc) && batch.token) {
                const int64_t n_tokens = cur->ne[0];

                memcpy(cur->data, batch.token, n_tokens*ggml_element_size(cur));
            }

            alloc_inp_tokens = true;
        }

        if (!alloc_inp_embd && strcmp(name, "inp_embd") == 0) {
            ggml_allocr_alloc(lctx.alloc, cur);