开启paddlelite支持在aarch64架构上编译不成功

环境

aarch64架构cpu 鲲鹏920 麒麟linux

问题说明：

想要编译一个在aarch64架构的上支持ppocr系列模型的fastdeploy-python。按文档，需要开启paddlelite支持。

按如下命令编译：
'''
git clone https://github.com/PaddlePaddle/FastDeploy.git
cd FastDeploy/python
export ENABLE_ORT_BACKEND=ON
export ENABLE_LITE_BACKEND=ON
export ENABLE_PADDLE_BACKEND=OFF
export ENABLE_OPENVINO_BACKEND=OFF
export ENABLE_VISION=ON
export ENABLE_TEXT=ON

python setup.py build
python setup.py bdist_wheel
'''
发现编译不通过。

初步排查有如下问题：

1. 链接paddlelite的so文件失败。看了下代码，在编译时，会下载lite-linux-arm64-20221209.tgz，然后其中lite-linux-arm64-20221209\lib里包含一个._libpaddle_full_api_shared.so文件，这个文件是苹果电脑的索引文件，会导致构建脚本误以为它是共享库：
   ’‘’
   AssertionError: patchelf /data/ai/fastdeploy/FastDeploy/python/.setuptools-cmake-build/third_libs/install/paddlelite/lib/._lib
   paddle_full_api_shared.so failed, the command: /data/ai/fastdeploy/FastDeploy/python/.setuptools-cmake-build/third_libs/patchelf/bin/patchelf --set-rpath '$ORIGIN:$ORIGIN/mklml/lib/' /data/ai/fastdeploy/FastDeploy/python/.setuptools-cmake-build/third_l
   ibs/install/paddlelite/lib/._libpaddle_full_api_shared.so
   ‘’‘
   这个问题是1.02引入的，之前是lite-linux-arm64-20220920.tgz

2. 我把paddlelite的.so文件替换掉，可以编译成功（1.07版本），但是现在运行程序时报：
   '''
   2024-11-14 17:26:17,794 - uvicorn.error - ERROR - Traceback (most recent call last):
   File "/root/miniconda3/envs/fastdeploy/lib/python3.10/site-packages/fastdeploy/c_lib_wrap.py", line 164, in
   from .libs.fastdeploy_main import *
   ImportError: /root/miniconda3/envs/fastdeploy/lib/python3.10/site-packages/fastdeploy/libs/libfastdeploy.so.1.0.7: undefined s
   ymbol: ZN6paddle8lite_api6Tensor11CopyFromCpuIlLNS0_10TargetTypeE1EEEvPKT
   '''

3. 由于还是未能成功运行。尝试用1.02版本使用lite后端运行ppocrv4，推理过程仍有错误。这部分错误没保存下来。

4. 最后发现文档里写的，ppocr在aarch64架构linux只支持paddlelite是不对的，实际上支持onnx_runtime，我直接用1.02版本开启onnx_runtime后端就成功运行了。

Start with python/setup.py and the PaddleLite download and library-install steps described in the report; inspect how files in the extracted lib directory are selected and patched on aarch64. Reproduce the build with ENABLE_LITE_BACKEND=ON, then check the generated wheel and c_lib_wrap.py import; done means PaddleLite builds and the resulting package imports and runs the reported PP-OCR case, or the documentation correctly identifies the supported backend.