Function llm_load_sparse_model_tensors

llama.cpp:3159–3390

}

static void llm_load_sparse_model_tensors(
        llama_model_loader & ml,
        llama_model & model,
        const llama_context_params * cparams,
        int main_gpu,
        long int vram_budget_bytes,
        bool reset_gpu_index,
        bool disable_ffn_split,
        bool use_mlock,
        llama_progress_callback progress_callback,
        void * progress_callback_user_data) {
    model.t_start_us = ggml_time_us();
    auto & ctx     = model.ctx;
    auto & hparams = model.hparams;

    size_t ctx_size;
    size_t mmapped_size;
    ml.calc_sizes(ctx_size, mmapped_size);
    LLAMA_LOG_INFO("%s: ggml ctx size = %7.2f MB\n", __func__, ctx_size/1024.0/1024.0);

    // create the ggml context
    {
        model.buf.resize(ctx_size);
        if (use_mlock) {
            model.mlock_buf.init   (model.buf.data);
            model.mlock_buf.grow_to(model.buf.size);
        }

        struct ggml_init_params params = {
            /*.mem_size   =*/ model.buf.size,
            /*.mem_buffer =*/ model.buf.data,
            /*.no_alloc   =*/ ml.use_mmap,
        };

        model.ctx = ggml_init(params);
        if (!model.ctx) {
            throw std::runtime_error(format("ggml_init() failed"));
        }
    }

    (void) main_gpu;

    enum ggml_backend_type llama_backend_offload = GGML_BACKEND_CPU;
    enum ggml_backend_type llama_backend_offload_split = GGML_BACKEND_CPU;

#ifdef GGML_USE_CUBLAS
    if (ggml_cublas_loaded()) {
        LLAMA_LOG_INFO("%s: using " GGML_CUDA_NAME " for GPU acceleration\n", __func__);
        ggml_cuda_set_main_device(main_gpu);

        llama_backend_offload = GGML_BACKEND_GPU;
        llama_backend_offload_split = GGML_BACKEND_GPU_SPLIT;
    }
#elif defined(GGML_USE_CLBLAST)
        LLAMA_LOG_INFO("%s: using OpenCL for GPU acceleration\n", __func__);
        llama_backend_offload = GGML_BACKEND_GPU;
        llama_backend_offload_split = GGML_BACKEND_GPU;
#endif