Android端推理LightGlue模型crash了

标题：尝试用Paddle-Lite推理开源的Lightglue模型，一调Run()函数就crash了
版本、预测库信息： 1）Paddle Lite 版本：v2.13rc 2）Host 环境： 3）运行设备环境：展锐T760(UMS9620W)/Android 4）预测后端信息：CPU/GPU
预测信息 1）预测 API：C++ 2）预测选项信息：arm64-v8a、4线程 3）预测库来源：官网下载
复现信息：如为报错，请提供剥离业务环境的可复现预测 demo，用于问题复现开源模型：https://github.com/fabio-sim/LightGlue-ONNX.git 模型转化路径：pytorch-->onnx-->nb x2paddle --framework=onnx \ --model=lightglue_1.onnx \ --save_dir=lightglue_1_paddlelite \ --to_lite=True \ --lite_valid_places=arm \ --lite_model_type=naive_buffer

问题描述：请详细描述您的问题，同步贴出报错信息、日志/代码关键片段代码：

    // -------------------------step 1: create predictor------------------------------ //
    paddle::lite_api::MobileConfig mb_config;
    mb_config.set_model_from_file(config->model_path);
    mb_config.set_threads(config->num_threads);
    //mb_config.set_power_mode(paddle::lite_api::LITE_POWER_HIGH);
    predictor = CreatePaddlePredictor<paddle::lite_api::MobileConfig>(mb_config);
    ...
    // -------------------------step 2: copy input data------------------------------ //
    std::unique_ptr<paddle::lite_api::Tensor> input_kpts0_tensor(std::move(predictor->GetInput(0)));
    input_kpts0_tensor->Resize({1, keypoints0.shape[2], 2});
    auto* input_kpts0_data = input_kpts0_tensor->mutable_data<float>();
    memcpy(input_kpts0_data, keypoints0.data, keypoints0.length * sizeof(float));

    std::unique_ptr<paddle::lite_api::Tensor> input_kpts1_tensor(std::move(predictor->GetInput(1)));
    input_kpts1_tensor->Resize({1, keypoints1.shape[2], 2});
    auto* input_kpts1_data = input_kpts1_tensor->mutable_data<float>();
    memcpy(input_kpts1_data, keypoints1.data, keypoints1.length * sizeof(float));

    std::unique_ptr<paddle::lite_api::Tensor> input_desc0_tensor(std::move(predictor->GetInput(2)));
    input_desc0_tensor->Resize({1, descriptors0.shape[2], descriptors0.shape[3]});
    auto* input_desc0_data = input_desc0_tensor->mutable_data<float>();
    memcpy(input_desc0_data, descriptors0.data, descriptors0.length * sizeof(float));

    std::unique_ptr<paddle::lite_api::Tensor> input_desc1_tensor(std::move(predictor->GetInput(3)));
    input_desc1_tensor->Resize({1, descriptors1.shape[2], descriptors1.shape[3]});
    auto* input_desc1_data = input_desc1_tensor->mutable_data<float>();
    memcpy(input_desc1_data, descriptors1.data, descriptors1.length * sizeof(float));

    // -------------------------step 3: run model------------------------------ //
    timespec begin, end;
    gettime(begin);
    predictor->Run();    // 这里crash了
    gettime(end);
    LOGD("Inference -- step 3: Paddle-Lite, %d threads, run lightglue complete, time: %d",
                this->config.num_threads, get_milliseconds(begin, end));

    // -------------------------step 4: parse output--------------------------- //
    std::vector<std::string> output_names = predictor->GetOutputNames();
    assert(output_names.size() == 1);
    ...

调用Run()函数就crash了，报错日志： 2024-04-30 14:59:17.576 18781-18781/com.abcde.pointglue A/libc: Fatal signal 11 (SIGSEGV), code 1 (SEGV_MAPERR), fault addr 0x70d5a85514 in tid 18781 (abcde.pointglue), pid 18781 (abcde.pointglue)

PaddlePaddle / Paddle-Lite

Android端推理LightGlue模型crash了 #10509