These days, I worked on running the MLPERF benchmark submitted by Nvidia. Firstly, I tried to not use nvidia-docker, but I failed. I don’t know what version of fairseq do you use in the docker image. It is totally different from the version I can find on github. Secondly, I use the docker you provided and ran it successfully. Then I uninstalled the original pytorch and reinstalled my own version which disabled CUDA P2P. However,
ImportError: /workspace/translation/fairseq/data/batch_C.cpython-36m-x86_64-linux-gnu.so: undefined symbol: _ZN2at5ErrorC1ENS_14SourceLocationERKNSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEE.
Could you please explain me about the fairseq you use? I can hardly solve this problem. Thanks!