Skip non-selected experts for mixtral and qwen2_moe #32429
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
@ArthurZucker, I encountered a conflict between torch.fx and a dynamic for loop in my implementation. I haven't found a concise solution for this issue yet. Could you help or just give up for this? |
|
Thanks for the info. I am not quite familiar with 'torch.fx'. Seems like there is a trade-off between enabling FX tracing and skipping experts. |
ArthurZucker
reviewed
Aug 8, 2024
There was a problem hiding this comment.
#31173 is related as well.
Yeah if fx's failing not 100% sure we want to do this, but we do needs benches of some sort regarding reduced IO, because for qwen with a lot of experts, it can be significatn
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Fixes #32283
This PR avoids redundant computation for some MoE models (mixtral and qwen2_moe).
The current implementation loops all the experts and inevitably loads experts' weight, which brings extra IO costs.
@ArthurZucker