I notice that README says: "Currently fairseq-py requires PyTorch version >= 0.4.0". But the latest pytorch version seems to be 0.3.1 ...
Hello, Could you tell me how to deal with it?
First, install the latest pytorch from source code according this:
https://github.com/pytorch/pytorch#from-source.
Second, install nccl2, if you want to train your own fairseq model [not nccl1!!!!!!!!!!!]:
http://docs.nvidia.com/deeplearning/sdk/nccl-install-guide/index.html#down
Finally, clone the latest fairseq-py and build the code:
https://github.com/facebookresearch/fairseq-py [the link of nccl in this page is redirected to nccl 1, do not install it]
thank you for your reply!
Most helpful comment
You should reinstall the pytorch from source code, because the lastest version[0.4 version] is not released yet.
First, install the latest pytorch from source code according this:
https://github.com/pytorch/pytorch#from-source.
Second, install nccl2, if you want to train your own fairseq model [not nccl1!!!!!!!!!!!]:
http://docs.nvidia.com/deeplearning/sdk/nccl-install-guide/index.html#down
Finally, clone the latest fairseq-py and build the code:
https://github.com/facebookresearch/fairseq-py [the link of nccl in this page is redirected to nccl 1, do not install it]