RL4LMs icon indicating copy to clipboard operation
RL4LMs copied to clipboard

Published 20 hours ago • verified publisher icon allenai

Any plans for Deepspeed/Accelerate integration?

Open Breakend opened this issue 3 years ago • 9 comments

Hi, great library! I'm wondering if you have any plans for deepspeed or accelerate integration to train larger models (e.g., GPT-J)?

Breakend avatar Oct 18 '22 07:10 Breakend

Hey @Breakend Thanks! We are working on the integration of HF Accelerate. Will let you know once it is out!

rajcscw avatar Oct 18 '22 07:10 rajcscw

Thanks! No pressure. Just curious is there any ETA?

boblee22 avatar Oct 18 '22 23:10 boblee22

@boblee22 Probably by end of this month :)

rajcscw avatar Oct 19 '22 20:10 rajcscw

Hi, thanks for the great library! Curious whether there's an updated eta on this issue?

Shikib avatar Dec 23 '22 01:12 Shikib

We are a bit delayed on this. But it is coming soon. Is there a particular LM that you would like to train (that needs multi node setup) ?

rajcscw avatar Dec 23 '22 11:12 rajcscw

I was trying to train blenderbot-3B, but it doesn't have to be that specific model. I'm interested in training models with 3-5B parameters, which I can currently do with deepspeed + huggingface. Thanks for prompt response!

Shikib avatar Dec 26 '22 19:12 Shikib

Any updates on this?

avacaondata avatar Mar 30 '23 08:03 avacaondata

We have a branch for this. https://github.com/allenai/RL4LMs/tree/add-accelerate-support But we are still testing it thoroughly before rolling it out.

rajcscw avatar Mar 31 '23 20:03 rajcscw

Perfect! Do you recommend me to try it out or should I wait until the release ?

avacaondata avatar Apr 12 '23 08:04 avacaondata