Method llama_gpu_split_loader

llama.cpp:2795–2814

    size_t vram_required = 0;

    llama_gpu_split_loader(const std::string & fname, bool use_mmap) : fname(fname), use_mmap(use_mmap) {
        GGML_ASSERT(use_mmap);

        idx_loader = new llama_model_loader(fname, use_mmap);
        GGUF_GET_KEY(idx_loader->ctx_gguf, vram_required, gguf_get_val_u64, GGUF_TYPE_UINT64, true, LLM_KV_NAMES[LLM_KV_SPLIT_VRAM_CAPACITY]);
        printf("loaded gpu_idx, vram_required: %ld\n", vram_required);

        n_tensors = idx_loader->n_tensors;

        // allocate memadata/data for mlp tensors
        // TODO: support allocating buffer for tensor data (when mmap is not used)
        size_t per_tensor_meta_size = GGML_PAD(sizeof(struct ggml_tensor), GGML_MEM_ALIGN) + GGML_OBJECT_SIZE;
        size_t tensor_meta_size = n_tensors * per_tensor_meta_size;
        struct ggml_init_params params = {
            /*.mem_size   =*/ tensor_meta_size,
            /*.mem_buffer =*/ nullptr,
            /*.no_alloc   =*/ true,
        };
        ctx_meta = ggml_init(params);
    }

    bool check_vram_allocable(size_t vram_budget) {
        return vram_budget >= vram_required;