I have changed NCCL to 'Off’in cmakelists and added USE_NCCL = False in setup.py,why it still make the nccl?How can I solve the problem?
This has been around for a little bit, but I have a similar issue on one particular system I am working on.
From the documentation, and looking through the source code it is not obvious which/how the setup script uses environmental variables (if at all), which makes it very confusing.
On this particular system nccl fails to compile, (its a cray system if that helps). I have set USE_NCCL=OFF as an environmental (I also tried USE_NCCL=0, USE_NCCL=False). I have also set this in tools/setup_helpers/nccl.py. Short of manually editing all mentions of NCCL in there a way to tell pytorch not to install NCCL, am I missing something obvious? Thanks in advance
Thank you for filing the issue about it. I’m proposing a fix (but use USE_DISTRIBUTED=OFF).
Best regards
Thomas