Method buffered_alloc

llama.cpp:3011–3027

    buffered_tensor_allocator(llama_model_loader &ml, ggml_context *ctx, const llama_hparams &hparams) : ml(ml), ctx(ctx), hparams(hparams) {}

    ggml_tensor * buffered_alloc(const std::string & name, const llm_tensor tensor_type, const std::vector<int64_t> & ne, const int i_layer) {
#if defined(GGML_USE_CUBLAS)
        tensor_offloading_levels level = get_offloading_level(tensor_type);
        if (level == TENSOR_NO_OFFLOAD || level == TENSOR_OFFLOAD_FFN) {
            return ml.create_tensor(ctx, name, ne, GGML_BACKEND_CPU);
        }
        // Alloc only metadata for GPU tensors
        bool no_alloc = ctx->no_alloc;
        ggml_set_no_alloc(ctx, true);
        ggml_tensor * meta_tensor = ml.create_tensor(ctx, name, ne, GGML_BACKEND_CPU);
        ggml_set_no_alloc(ctx, no_alloc);
        alloc_queues[level].push_back(std::make_tuple(i_layer, tensor_type, meta_tensor));
        return meta_tensor;
#else
        return ml.create_tensor(ctx, name, ne, GGML_BACKEND_CPU);
#endif
    }

    bool offload_tensor(ggml_tensor * meta_tensor) {
        size_t tensor_data_size = ggml_nbytes(meta_tensor);