Closed jambo6 closed 5 months ago
Looks like I have some old ver of triton somehow that does not have this commit
https://github.com/openai/triton/pull/1712/files
not sure how as we are on 2.1.0 but anyway
In case anyone has this issue, seems to be related to this issue and using a pytorch docker image which does not have the correct version of the code
Have you been able to fix this error?
Nvm- for future reference:
you may have triton==2.1.0 installed from your docker container, but it may not match the latest triton version (they likely updated the pkg). So make sure to add pip install --force-reinstall triton==2.1.0
. This will bring your triton version 2.1.0 up to date with the actual version on the openai repo
I get the following error when I try to run bwds pass with
moe_expert_model_parallelism=False
. E.g. if I runmoe_test.py
it fails with this error.My versions