Closed HeegyuKim closed 1 year ago
thanks for the report ill change the llama rotary type default to complex in version 0.0.20
If you wonder why are there three types of rotary embeddings in easydel, I have to tell you that in the case of using flash attention and training the lm2 and open are faster than complex for rotary embedding but complex if slower and more accurate and I'm still trying to find another way for that
and if you want to have better using or hosting and faster generation I recommend you to use JAXServer if you have any other problems please let me know
I added from_pt
to the llama models you don't need to change that yourself anymore :)
read NOTE.md if you do like to see that
In your code, default rotary type is 'lm2' but generation output is strange.
model is
meta-llama/Llama-2-7b-hf
rope type = 'lm2'
rope type = 'complex'
This is my test code. I changed parameter names of llama model to use huggingface
from_pt=True
When I used converted weights using converter.py, generation output was also strange.