Function llm_build_kv_store

llama.cpp:4383–4418

}

static std::pair<ggml_tensor*, ggml_tensor*> llm_build_kv_store(
        struct ggml_context * ctx,
        const llama_hparams & hparams,
       const llama_kv_cache & kv,
         struct ggml_cgraph * graph,
         struct ggml_tensor * k_cur,
         struct ggml_tensor * v_cur,
                    int64_t   n_ctx,
                    int32_t   n_tokens,
                    int32_t   kv_head,
         const llm_build_cb & cb,
                    int64_t   il) {
    const int64_t n_embd_gqa = hparams.n_embd_gqa();

    // compute the transposed [n_tokens, n_embd] V matrix
    struct ggml_tensor * v_cur_t = ggml_transpose(ctx, ggml_reshape_2d(ctx, v_cur, n_embd_gqa, n_tokens));
    //struct ggml_tensor * v_cur_t = ggml_transpose(ctx, v_cur); // TODO: reshape above is likely not needed
    cb(v_cur_t, "v_cur_t", il);

    struct ggml_tensor * k_cache_view = ggml_view_1d(ctx, kv.k, n_tokens*n_embd_gqa,
            (ggml_element_size(kv.k)*n_embd_gqa)*(il*n_ctx + kv_head));
    cb(k_cache_view, "k_cache_view", il);

    struct ggml_tensor * v_cache_view = ggml_view_2d(ctx, kv.v, n_tokens, n_embd_gqa,
            (   n_ctx)*ggml_element_size(kv.v),
            (il*n_ctx)*ggml_element_size(kv.v)*n_embd_gqa + kv_head*ggml_element_size(kv.v));
    cb(v_cache_view, "v_cache_view", il);

    // important: storing RoPE-ed version of K in the KV cache!
    ggml_tensor * k_cpy = ggml_cpy(ctx, k_cur,   k_cache_view);
    ggml_tensor * v_cpy = ggml_cpy(ctx, v_cur_t, v_cache_view);
    //ggml_build_forward_expand(graph, ggml_cpy(ctx, k_cur,   k_cache_view));
    //ggml_build_forward_expand(graph, ggml_cpy(ctx, v_cur_t, v_cache_view));
    
    return {k_cpy, v_cpy};
}

static struct ggml_tensor * llm_build_norm(
        struct ggml_context * ctx,