Method llm_build_plamo3

subprojects/llama.cpp/src/models/plamo3.cpp:4–124

template <bool iswa>
llm_build_plamo3<iswa>::llm_build_plamo3(const llama_model & model, const llm_graph_params & params) :
    llm_graph_context(params) {
    const int64_t head_dim_q = hparams.n_embd_head_k;
    const int64_t head_dim_v = hparams.n_embd_head_v;

    ggml_tensor * cur;
    ggml_tensor * inpL = build_inp_embd(model.tok_embd);
    ggml_tensor * inp_pos = build_inp_pos();

    using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
    inp_attn_type * inp_attn = nullptr;

    if constexpr (iswa) {
        inp_attn = build_attn_inp_kv_iswa();
    } else {
        inp_attn = build_attn_inp_kv();
    }

    ggml_tensor * inp_out_ids = build_inp_out_ids();

    for (int il = 0; il < n_layer; ++il) {
        ggml_tensor * residual = inpL;

        float freq_base_l  = 0.0f;
        float freq_scale_l = 0.0f;
        if constexpr (iswa) {
            freq_base_l  = model.get_rope_freq_base (cparams, il);
            freq_scale_l = model.get_rope_freq_scale(cparams, il);
        } else {
            freq_base_l  = freq_base;
            freq_scale_l = freq_scale;
        }

        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
        cb(cur, "attn_norm", il);

        ggml_tensor * qkv = build_lora_mm(model.layers[il].wqkv, cur);
        cb(cur, "wqkv", il);

        const int32_t n_head    = hparams.n_head(il);
        const int32_t n_head_kv = hparams.n_head_kv(il);

        const int64_t q_offset = 0;
        const int64_t k_offset = head_dim_q * n_head;
        const int64_t v_offset = k_offset + head_dim_q * n_head_kv;

        ggml_tensor * Qcur = ggml_view_3d(ctx0, qkv, head_dim_q, n_head, n_tokens,
                head_dim_q * sizeof(float), qkv->nb[1], q_offset * ggml_element_size(qkv));
        ggml_tensor * Kcur = ggml_view_3d(ctx0, qkv, head_dim_q, n_head_kv, n_tokens,
                head_dim_q * sizeof(float), qkv->nb[1], k_offset * ggml_element_size(qkv));
        ggml_tensor * Vcur = ggml_view_3d(ctx0, qkv, head_dim_v, n_head_kv, n_tokens,
                head_dim_v * sizeof(float), qkv->nb[1], v_offset * ggml_element_size(qkv));

        cb(Qcur, "Qcur", il);
        cb(Kcur, "Kcur", il);
        cb(Vcur, "Vcur", il);

        Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);