model_server: ovms has a segment error when loading a shared library using the AVX512 instruction set using cpu_extension

System information (version)

  • OpenVINO => 2021.3
  • Operating System / Platform => linux
  • Compiler => gcc 8
  • OpenVINO model server 2021.3 docker image
#./ovms --model_path /data/lat_open --model_name lat_open --port 9000 --cpu_extension /data/libextension.so --plugin_config '{"CPU_THROUGHPUT_STREAMS": "12"}' --shape auto
[2021-08-13 09:40:12.520][988][serving][info][server.cpp:105] OpenVINO Model Server 2021.3.c9ba5cfa
[2021-08-13 09:40:12.520][988][serving][info][server.cpp:106] OpenVINO backend 2021.3.394
[2021-08-13 09:40:12.520][988][serving][info][model.cpp:38] Getting model from /data/lat_open
[2021-08-13 09:40:12.520][988][serving][info][model.cpp:45] Model downloaded to /data/lat_open
[2021-08-13 09:40:12.520][988][serving][info][model.cpp:145] Will add model: lat_open; version: 1 ...
[2021-08-13 09:40:12.520][988][serving][info][modelinstance.cpp:495] Loading model: lat_open, version: 1, from path: /data/lat_open/1, with target device: CPU ...
[2021-08-13 09:40:12.520][988][serving][info][modelinstance.cpp:499] Some inputs shapes for model lat_open are set to auto
[2021-08-13 09:40:12.520][988][serving][info][modelversionstatus.hpp:155] STATUS CHANGE: Version 1 of model lat_open status change. New status: ( "state": "START", "error_code": "OK" )
[2021-08-13 09:40:12.520][988][serving][info][modelversionstatus.hpp:155] STATUS CHANGE: Version 1 of model lat_open status change. New status: ( "state": "LOADING", "error_code": "OK" )
[2021-08-13 09:40:12.521][988][serving][info][modelinstance.cpp:249] Loading custom CPU extension from /data/libextension.so
[2021-08-13 09:40:12.523][988][serving][info][modelinstance.cpp:252] Custom CPU extention loaded. Adding it.
Segmentation fault (core dumped)

if i do not avx512 instruction , it will be ok.

how to fix it, thank you very much.

About this issue

  • Original URL
  • State: closed
  • Created 3 years ago
  • Comments: 21

Most upvoted comments

@atobiszei can you reproduce it now?