PyTorch自定义算子及TensorRT插件demo

1. 编译torch自定义算子
   • cd cuda_op
   • python setup.py build_ext --inplace
   • cd /data/sunkx/workspace/TensorRT-plugin-demo
2. 模型测试、导出onnx
   • python model_symbolic.py
3. TensorRT插件编译
   • export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/data/sunkx/TensorRT-10.4.0.26/lib
   • export PATH=$PATH:/usr/local/cuda/bin
   • cd plugin
   • mkdir build
   • cd build
   • cmake ..
   • make -j
   • cd /data/sunkx/workspace/TensorRT-plugin-demo
4. 模型转换
   • /data/sunkx/TensorRT-10.4.0.26/bin/trtexec --onnx=./model_symbolic.onnx --plugins=./plugin/build/libcenterpool_plugin.so --saveEngine=model.engine --device=2
5. 模型推理
   • export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/data/sunkx/workspace/TensorRT-plugin-demo/plugin/build
   • cd infer
   • mkdir build
   • cd build
   • cmake ..
   • make -j
   • ./infer ../../model.engine