TensorRT Triton Inference Server: 版本 error魔术标记不匹配 ， NGC使用

魔术标记不匹配错误Serialization assertion magicTagRead == kMAGIC_TAG failed.Magic tag does not match

• 原因：
  转换和推理使用的镜像的标签是相同的，但是转换的镜像中pip list得到trt版本为8.6.0，但是推理环境中 root@f2c810ba3976:/# /usr/src/tensorrt/bin/trtexec ， 得到trtexec版本为8.5.3

• 解决
  可在官方文档中寻找正确版本的镜像地址https://catalog.ngc.nvidia.com/orgs/nvidia/containers/tritonserver/tags
  版本说名如下
  https://docs.nvidia.com/deeplearning/triton-inference-server/release-notes/rel-23-04.html#rel-23-04

• 小版本不一致
  The engine plan file is not compatible with this version of TensorRT, expecting library version 8.6.1.2 got 8.6.0.12, please rebuild.
  解决： rebuild

pytorch NGC

• https://catalog.ngc.nvidia.com/orgs/nvidia/containers/pytorch/tags
• sudo docker run -it --shm-size 8gb --rm --gpus=all -p 8603:8000 -v ${PWD}:/test nvcr.io/nvidia/pytorch:23.05-py3 bash

torch NGC

• https://catalog.ngc.nvidia.com/orgs/nvidia/containers/torch/tags

cuda NGC

• https://catalog.ngc.nvidia.com/orgs/nvidia/containers/cuda/tags

tensorrt NGC

• https://catalog.ngc.nvidia.com/orgs/nvidia/containers/tensorrt/tags