Conversation
open_lm/params.py
Outdated
| 'Allows int8 training/inference, etc.' | ||
| ) | ||
| parser.add_argument( | ||
| "--xformers-rotary", |
There was a problem hiding this comment.
Can we call this flag something like "--old-xformers-rotary" so that it still makes sense when they fix the bug?
There was a problem hiding this comment.
hmm so when this gets fixed upstream this flag will use the fix so it won’t be “old” i guess we want this to always use the old implementation and will make changes accordingly
There was a problem hiding this comment.
hmm that's a good point..
There was a problem hiding this comment.
Just to check: can we set up things so that when the fix happens upstream, we won't need a flag in OpenLM? I think in general the fewer flags a user needs to pass in for regular use, the better.
There was a problem hiding this comment.
that’s how things are now but i think we do need to support an “old” flag here considering that we have already trained models with the head rotary
|
confirmed that this looks good still with most up to date changes |
|
Amazing! |
Green is after fix
