Add OLMoE #32406
Add OLMoE #32406
Conversation
It's ready! :) I will update the README & double-check the slow tests once the model is released if that's fine! |
|
@ArthurZucker would be great if we could get it reviewed soon 😇 |
|
We'll release the model on Tuesday, would be amazing if we could have this approved by then! |
|
Oups sorry reviewing now |
| @@ -0,0 +1,281 @@ | |||
| # Licensed under the Apache License, Version 2.0 (the "License"); | |||
There was a problem hiding this comment.
date is missing here!
There was a problem hiding this comment.
Is that mandatory? Maybe the entire header can just be removed? Seems redundant to have this header for every file when the license is clear from the repo..
There was a problem hiding this comment.
😄 yeah it is redundant but the licence AFAIK but it's a nit don't worry
| return final_hidden_states, router_logits | ||
|
|
||
|
|
||
| class OlmoeDecoderLayer(nn.Module): |
There was a problem hiding this comment.
Different from others cuz we have no shared expert
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
|
COuld you make sure you rebased and CIs are green? Will review again after that! |
|
Great, all passing! 🙌 |
|
Thanks for your contribution! 🔥 |
|
Is it a problem on my end if I notice that a higher native batch size during training results in higher losses? (With DeepSpeed)
Purple line is higher native bs FFT, orange line is gradient accumulation instead of a higher native batch size. I'm guessing expert routing parallelism is maybe not properly handled for bs>1? |
|
cc @Muennighoff would have a lot more clues than me 😉 |
Hm not sure about this - are other models the same for you in both scenarios? How about other MoEs? |
What does this PR do?
Before submitting
Pull Request section?
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
The model will be released in ~1 week - can we already review this so that we can merge right upon release?