Method build_inp_cross_embd

subprojects/llama.cpp/src/llama-graph.cpp:1521–1543

}

ggml_tensor * llm_graph_context::build_inp_cross_embd() const {
    auto inp = std::make_unique<llm_graph_input_cross_embd>(cross);

    auto & cur = inp->cross_embd;

    // if we have the output embeddings from the encoder, use them directly
    // TODO: needs more work to be correct, for now just use the tensor shape
    //if (cross->t_embd) {
    //    cur = ggml_view_tensor(ctx0, cross->t_embd);

    //    return cur;
    //}

    const auto n_embd = !cross->v_embd.empty() ? cross->n_embd : hparams.n_embd_inp();
    const auto n_enc  = !cross->v_embd.empty() ? cross->n_enc  : hparams.n_ctx_train;

    cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_enc);
    ggml_set_input(cur);

    res->add_input(std::move(inp));

    return cur;
}

ggml_tensor * llm_graph_context::build_inp_pos_bucket_enc() const {
    auto inp = std::make_unique<llm_graph_input_pos_bucket>(hparams);