Function llm_load_gpu_split

llama.cpp:3142–3157

}

static size_t llm_load_gpu_split(llama_model_loader & ml, llama_model & model, bool no_cache, bool no_offload) {
#if defined (GGML_USE_CUBLAS)
    if (!ggml_cublas_loaded()) {
        throw std::runtime_error(format("cannot offload to GPU: " GGML_CUDA_NAME " not loaded"));
    }
    if (!no_offload && !llm_load_gpu_split_with_budget(ml, model, vram_budget_bytes, no_cache)) {
        LLAMA_LOG_ERROR("%s: error: failed to generate gpu split, an empty one will be used\n", __func__);
    }
#endif

    // Apply GPU index and split FFNs to GPU
    size_t ffn_offloaded_bytes = llama_model_offload_ffn_split(&model);
    LLAMA_LOG_INFO("%s: offloaded %.2f MiB of FFN weights to GPU\n", __func__, ffn_offloaded_bytes / 1024.0 / 1024.0);

    return ffn_offloaded_bytes;
}

static void llm_load_sparse_model_tensors(
        llama_model_loader & ml,